Australia’s National

Science Agency

Artificial intelligence
for science
Adoption trends and future development pathways

November 2022
Citation Acknowledgements
Hajkowicz S, Naughtin C, Sanderson C, Schleiger E, The authors of this study would like to thank the many
Karimi S, Bratanova A, Bednarz T (2022). Artificial artificial intelligence experts within CSIRO and beyond
intelligence for science – Adoption trends and future who generously shared their time and knowledge through
development pathways. CSIRO Data61, Brisbane, Australia. consultative interviews. The authors would also like to
thank staff at The Lens (lens.org) who provided support,
advice and guidance on our bibliometric analysis.
Copyright
© Commonwealth Scientific and Industrial Research
Organisation 2022. To the extent permitted by law, all Units and Notation
rights are reserved, and no part of this publication covered Unless otherwise specified, dollar values in this report
by copyright may be reproduced or copied in any form or are given in Australian dollars in current prices and this
by any means except with the written permission of CSIRO. report uses the International System of Units (SI).

Important disclaimer Accessibility

CSIRO advises that the information contained in this CSIRO is committed to providing web accessible content
publication comprises general statements based on wherever possible. If you are having difficulties with
scientific research. The reader is advised and needs to accessing this document please contact csiro.au/contact.
be aware that such information may be incomplete or
unable to be used in any specific situation. No reliance
or actions must therefore be made on that information
without seeking prior expert professional, scientific
and technical advice. To the extent permitted by
law, CSIRO (including its employees and consultants)
excludes all liability to any person for any consequences,
including but not limited to all losses, damages, costs,
expenses and any other compensation, arising directly
or indirectly from using this publication (in part or in
whole) and any information or material contained in it.

Cover image: 3D SLAM (Simultaneous Localisation and Mapping)

mapping software Wildcat creates a detailed scan of Fish Lane
in Brisbane to act as a map for an autonomous robot.
Executive summary
Key findings

1 2 3
We are amidst the largest surge The future impacts of AI We identify six future
in history in the development range from enablement to development pathways for
and adoption of artificial transformation. Enabling research organisations seeking
intelligence (AI) for scientific means improving the speed, to upgrade AI capability for
research in all disciplines safety, cost‑effectiveness and the future so they can harness
of natural science, physical quality of research. It’s also the benefits and mitigate
science, social science and the possible that AI could transform the risks of AI technologies.
arts and humanities. This will knowledge discovery, allowing We provide detailed guidance
impact the research sector, scientists to solve problems on each pathway.
research organisations and hitherto beyond reach.
individual research careers.

In the wetlands of Kakadu, rangers

are using AI and Indigenous
Knowledge to care for country.
Together with the Kakadu Rangers
and Microsoft, we developed the
Healthy Country Dashboard.
Image: Michael Douglas

i
Why we wrote this report The AI approach improved productivity by 600 times
[5]. In another case, CSIRO researchers used machine
This report aims to inform researchers and research learning to find disease genes. This approach processes
organisations within the spheres of government, 10 million genomic variants in 15 hours compared to
industry, community and academia seeking to develop traditional approaches which are estimated to take
improved AI capabilities. The report is focused on the 100,000 years [6]. There are many more such examples.
use of AI for science, and it describes AI adoption trends
in the physical, natural and social science fields. Using a AI development can also be challenging. Whilst AI
bibliometric analysis of peer-reviewed publishing trends is a powerful technology, the pathway to successful
over 63 years (1960–2022), the report demonstrates a application within research fields can be challenging
surge in AI adoption across all fields over the past several and holds some level of risk. For example, a review of
years. The report examines future development pathways 62 machine-learning models developed for COVID-19
and explores implications for science organisations. diagnosis from chest scans found that none were
sufficiently reliable for clinical application [7].

Artificial intelligence
adoption trends Future development pathways
Software and hardware upgrades. Research organisations
There is a recent surge in publications of AI adoption
seeking to upgrade AI capability can take advantage
by scientists. By September 2022 approximately
of the continuously improving AI toolkit. Purpose-built
5.7% of all peer-reviewed research worldwide was
processors designed for machine learning are speeding
on the topic of AI. This is up from 3.1% in 2017 and
up computations. Quantum computing could lead to
1.2% in 2000. In 2020, before research had migrated
transformative leaps in computational power. Platforms
to COVID-19 topics, five (of seven) of the most
such as PyTorch, TensorFlow, Microsoft Azure and Amazon
influential research papers on Google Scholar (based
Web Services along with code-free AI software tools are
on citation metrics) were on the topic of AI [1].
making it easier for scientists to develop and apply AI.
AI is being used in increasingly diverse research domains.
The quest for better data. The era of ‘big data’ may
In the field of computer science, 30% of all peer-reviewed
be transitioning into the era of ‘better data’. Recent
papers are now on the topic of AI. However, the use of
breakthroughs in machine learning, capable of
AI extends well beyond computer science. Practically all
supporting mission-critical, real-world applications,
fields of natural, physical and social science and the arts
have been achieved using smaller datasets that are
and humanities are rapidly absorbing AI technology.
well curated, fit-for-purpose and provenance assured.
Public and private sector AI research and development Targeted investment in quality datasets will allow
investment is rising. The average advanced-economy spend research organisations to develop AI capabilities.
on research and development (R&D) has risen from 2.1%
Education, training and capability uplift. There has
of GDP in 2000 to 2.5% in 2019 [2]. Within the R&D spend,
been an explosion of university and technical-college
resources devoted to AI are increasing. Since 2017 over
training degrees and courses on AI at undergraduate
700 AI policy and strategy initiatives have been developed
and postgraduate levels over the past 5 years.
across 60 countries and territorial jurisdictions [3, 4].
The number of students enrolling in these courses is
There is evidence of enhanced productivity through case rising. There’s been an increase in the diversity of fee-
studies. In this report we identify and review numerous based, or free-of-charge, professional development
case studies where AI has improved the efficiency and education and training for scientists wanting to upskill
effectiveness of research. For example, a CSIRO team in AI. Research organisations can take advantage of
developed an automated robotic system that tests wide‑ranging training and education resources to recruit
12,000 solar cells in 24 hours. Previously the research AI talent and uplift capabilities of existing staff.
team could manually test up to 20 solar cells per day.

ii Artificial intelligence for science

Our robots can perform multi‑agent
navigation to effectively explore
unknown environments.

Toward human-centric artificial intelligence. Conclusion

In the vast majority of cases, AI will be augmenting
not replacing the human scientist. Research This report demonstrates the widespread and rapid
organisations need to find ways via which humans diffusion of AI technology across all fields of natural,
work harmoniously with AI-enabled systems. Issues physical and social science, and the arts and humanities.
of trust, transparency and reliability will be important The current surge in uptake far exceeds the breadth and
for scientists and reviewers working on AI systems. depth of AI’s 2 historical surges. There are no signs of a
slowdown. Recent developments in AI are likely to have
Improving workforce diversity. The AI research a lasting impact on how humans discover knowledge
workforce lacks gender, ethnic and cultural diversity and solve problems in all fields of science and research.
which limits the quality of outcomes. Improving the However, the pathway to AI adoption and capability uplift
gender, ethnic and cultural diversity of the AI research is challenging. Researchers are likely to experience both
workforce will lead to better science outcomes. success and failure as they develop AI systems within their
domains of expertise. This report identifies pathways
Meeting societal expectations and regulations for
for organisations engaged in science and research
ethical AI. Research organisations will be challenged
seeking to uplift their AI capability into the future.
to develop capabilities, technologies and cultures that
deliver increasingly ethical AI. Societal expectations
for ethical AI are rising. Voluntary principles for
ethical AI may increasingly become regulations
and laws. For example, AI-specific legislation has
been proposed in Europe and the United States.

iii
Contents
1 Introduction..................................................................................................................................................1
2 Artificial intelligence – Why now?.....................................................................................................3
2.1 A turbulent history ........................................................................................................................................................3
2.2 Why is now different?....................................................................................................................................................4

3 The science sector....................................................................................................................................7

3.1 What is the science sector?...........................................................................................................................................7
3.2 The global science sector..............................................................................................................................................7
3.3 The Australian science sector......................................................................................................................................10
3.4 Productivity decline.....................................................................................................................................................14

4 Artificial intelligence and knowledge discovery....................................................................... 17

4.1 Enablement or transformation?..................................................................................................................................17
4.2 Case studies – Artificial intelligence applications for science..................................................................................18

5 Science domain adoption trends................................................................................................... 23

5.1 Data sources and methods (bibliometric analysis)....................................................................................................23
5.2 Publishing intensity and volumes – all research fields..............................................................................................26
5.3 Adoption trends in application domains ...................................................................................................................26
5.4 Artificial intelligence technology diffusion trends....................................................................................................32
5.5 Trends in artificial intelligence technologies.............................................................................................................34

6 Future development pathways ...................................................................................................... 37

6.1 Software, hardware and open access resources........................................................................................................37
6.2 The quest for better data.............................................................................................................................................38
6.3 Education, training and capability uplift....................................................................................................................39
6.4 Towards collaborative artificial intelligence..............................................................................................................42
6.5 Artificial intelligence workforce diversity................................................................................................................. 44
6.6 The rise of ethical expectations and regulations......................................................................................................45

7 Conclusion................................................................................................................................................. 47
8 References.................................................................................................................................................48
Appendix A – Science occupations.....................................................................................................56
Appendix B – Artificial intelligence phrases.................................................................................... 57

iv Artificial intelligence for science

Figures
Figure 1. Alan Turing – Can machines think? .........................................................................................................................3

Figure 2. Private investment in artificial intelligence companies worldwide (billions of US dollars).................................5

Figure 3. Increasing spending on R&D within the OECD. ......................................................................................................8

Figure 4. Share of global peer-reviewed publishing by research field in 2022. ...................................................................9

Figure 5. Worldwide peer-reviewed scientific publishing in all fields for 1960–2021. .....................................................10

Figure 6. The number of Australian scientists by industry grouping..................................................................................11

Figure 7. The number of Australian scientists by field of science. ......................................................................................11

Figure 8. Comparative levels of specialisation in Australian science and research during 2012–2021............................13

Figure 9. Average (mean) annual OECD-country multi-factor productivity growth..........................................................14

Figure 10. Identifying artificial intelligence scholarly works from 1960 to 2022...............................................................24

Figure 11. Peer-reviewed research publications on artificial intelligence..........................................................................27

Figure 12. Annual change in peer-reviewed AI-related publications (%) over time...........................................................27

Figure 13. Artificial intelligence publishing intensity by main research domains..............................................................29

Figure 14. Artificial intelligence publishing intensity in the physical sciences (1–5 fields)...............................................30

Figure 15. Artificial intelligence publishing intensity in the physical sciences (6–10 fields).............................................30

Figure 16. Artificial intelligence publishing intensity in the health sciences......................................................................31

Figure 17. Artificial intelligence publishing intensity in the life sciences...........................................................................31

Figure 18. Artificial intelligence publishing intensity in the social sciences and humanities...........................................32

Figure 19. Concentration of artificial intelligence publishing across research fields........................................................33

Figure 20. Diffusion of artificial intelligence technology into research fields...................................................................33

Figure 21. The six fastest-growth artificial intelligence technologies over past 20 years.................................................34

Figure 22. Patent citation counts for the top 20 artificial intelligence phrases for 2017–2021........................................35

v
Testing a bionic eye. In partnership with CSIRO
spinout Bionic Vision Technologies, we helped
develop a retinal implant along with a vision
processing and software library to restore sight to
a degree in people with profound vision loss due
to Retinitis Pigmentosa or Macular Degeneration.

vi Artificial intelligence for science

1 Introduction
We are currently amid the largest surge, arguably ‘boom’, in the application
and development of artificial intelligence (AI) for scientific research in
history. Scholarly publications, patents, education, training, salaries,
research activity and investment are increasing at unprecedented rates.
We may now be on the steepest part of the adoption and development
curve. This is happening across the entire economy. Practically all industry
sectors, advanced economies, professions and world regions are seeing
rapid uptake of AI. The science sector is no exception. There’s a worldwide
competitive race, and collaborative movement, to develop AI capability [8].
Many scientists, and science organisations, are aiming to uplift AI capability.

We are seeing steep rates of adoption in science because, Other research institutes examining the impact of AI
in many well-publicised cases, AI is improving the on science and exploring related issues include:
speed, cost-effectiveness, quality and safety of scientific
research. In some cases, the benefits from using AI are • The Alan Turing Institute. The institute was
transformational; scientists have been able to solve awarded £38.8 million ($70.6 million) in 2018
problems hitherto beyond reach. There is a hope that for a 5-year research program on ‘AI and Data
AI can provide a much-needed productivity boost for Science for Science, Engineering, Health and
science. AI may help scientists address humanity’s greatest Government’. This applied‑research program
challenges such as climate change, pollution, resource aims to understand and accelerate productive
scarcity and infectious diseases. However, not all AI projects application of AI within these sectors [9].
have met the expectations of scientists. Sometimes AI • The ‘Artificial Intelligence and Augmented
projects can be complex, costly, time-consuming and Intelligence for Automated Investigations for
labour-intensive with limited results. The pathway to Scientific Discovery’ (AI4SD) program aims to
AI enablement, which most science organisations have explore and demonstrate how AI technologies can
embarked upon, is both rewarding and challenging. boost discovery in all fields of research [10]. It is
funded by the United Kingdom Engineering and
This report has been prepared to help science managers, Physical Sciences Research Council. A recent AI4SD
science organisations and investors understand plausible conference was held at Chilworth Manor in the United
development pathways for AI. Our aim is to describe Kingdom (and online) during 1–3 March 2022.
how AI has changed science and what the future may
• The Organisation for Economic Co-operation and
hold. We hope this will help science-sector workers
Development (OECD). Under the broader umbrella of
make informed decisions about how they prepare for an
the AI Policy Observatory, the OECD held a conference
AI‑enabled future. Such decisions may be about investment,
on ‘artificial intelligence and the future of science’
divestment, capability uplift, education, training and
from 29 October to 5 November 2021. The conference
organisational design. We think most of the world’s
examined the science productivity slump and the extent
science and research organisations are currently working
to which AI may provide a solution [11]. At this conference
through these, and related, issues as they seek to harness
experts from across the globe presented data showing
the opportunities and mitigate the risks of AI technology.
declining productivity in the science sector and discussed
improvements possible via AI technology [12, 13].

1
• The Stanford University Human-Centered AI Institute. • The 20-year community roadmap for AI research in
This institute prepares the AI index and associated the United States [18]. This document, and associated
reports, which are regularly updated to provide a program of activity, is concerned with AI capability uplift
comprehensive dataset on AI technology uptake and in the United States out to the year 2040. It identifies
adoption across all fields of science, industry and research priorities in the areas of (a) integrated
society [14]. This institute also has a lead role in the intelligence, (b) meaningful interaction, and (c) self‑aware
‘One Hundred Year Study of Artificial Intelligence’ learning. The report makes recommendations about
(AI100), which provides an ‘insider’ perspective hardware and software resources, training and
about the past, present and future trajectory of AI education, ethics, policy, workforce transitions and
development. The most recent AI100 report was mission-led research for AI, amongst other matters [18].
published in 2021 and is titled ‘Gathering Strength,
Gathering Storms’ [15], indicating enhanced AI capability This report contributes to the understanding about how
but coming with increased risks and complexity. AI will enable, and potentially transform, science from a
global and Australian perspective. Our report opens with
• The Argonne Laboratory AI for Science Project.
a brief history of AI and what makes now, the current
The Argonne Laboratories at Oak Ridge and Berkeley
boom cycle, different from the past. We then describe
hosted ‘town hall’ meetings for over 1,000 scientists
the global and Australian science sectors, highlighting
during July to October 2019 about the use of AI, big
science’s productivity slump which AI can potentially
data and high-performance computing. The findings
help solve. We next present a bibliometric analysis of
were captured in the ‘AI for Science’ report written by
AI adoption across all science domains and patterns
over 80 authors. This report provides a detailed account
of AI science and technology development. Lastly, the
about the state of the art, grand challenges, advances
report explores AI development pathways in science
over the next decade, accelerating development
over the coming decade and examines the strategic
and expected outcomes from AI application in
implications for scientists and science organisations
nine major fields of scientific research [16].
aiming to uplift capability for an AI-enabled future.
• The University of Adelaide Australian Institute for
Machine Learning (AIML). Along with the Australian
Strategic Policy Institute, the AIML recently published
a report titled ‘Artificial intelligence: Your questions
answered’ [17], which examines issues of development,
adoption and adaptation to AI technologies in
Australia. It also examines issues of sovereign
capability and why Australian industry often cannot
buy AI ‘off-the-shelf’. The AIML actively monitors
and examines issues relating to AI in Australia.

2 Artificial intelligence for science

2 Artificial intelligence –
Why now?
2.1 A turbulent history
There are many detailed accounts of AI’s history. In this
section we draw upon key publications [19-23] to provide a
brief summary of AI’s historic development to contextualise
our subsequent analysis of its current status and future
development pathways. We emphasise that AI isn’t new
to science; it has been on a long and turbulent journey
with interest waxing and waning through history.

AI is a field of science that has been widely considered

done-and-dusted a few times. However, AI has shown a
strong ability to bounce back and re-establish scientific
prominence. It is hard, if not impossible, to identify a start
date for AI research. Scientists were publishing on concepts
related to AI in the 1930s and 1940s. For example, Walter
Pitts and Warren McCullough [24] published a paper in
1943 about how artificial neurons can perform logical
functions. One of their students, Marvin Minsky, later
developed the ‘stochastic neural analog reinforcement
calculator’ [25] which has evolved into AI neural networks
used widely in machine learning (deep learning) today. Figure 1. Alan Turing – Can machines think?
Data source: Sketch by Natata on Shutterstock.
One of the pivotal papers in AI was written by Alan Turing
(Figure 1) and published in October 1950 [26]. This paper
opens with the words ‘I propose to consider the question,
can machines think?’. It lays out the future challenges for by mathematician James Lighthill. The report was highly
AI to solve. Turing’s question is still being asked [27] but critical of AI’s failure to achieve its ‘grandiose objectives’
remains unanswered. Turing died in 1954. The field of AI got [29]. The funding agencies mostly agreed with the view
its name at the Dartmouth Workshop of 1956. Organised that AI had over-promised and under-delivered. The flow
by John McCarthy, Marvin Minsky, Nathaniel Rochester and of resources for AI research was reduced to a trickle [23].
Claude Shannon, this meeting brought together leading
Despite the setbacks, the 1980s saw a return to boom times
AI experts of the time [28]. Workshop attendees agreed to
with the rise of expert systems and connectionism – an
adopt ‘artificial intelligence’ as the name of their emerging
approach in the cognitive sciences which explains mental
research field. Naming AI helped connect a related set of
phenomena using artificial neural networks. The Japanese
technologies, concepts and theories. It helped formalise
Government began aggressively funding AI through
and establish an identity for a new field of science.
the fifth-generation computer project [30]. The United
Investment and activity in AI escalated during the 1950s Kingdom and United States governments were soon to
and, more so, in the 1960s. Significant advances occurred follow, again injecting substantial funds into a range of AI
in the fields of natural language processing, automated research initiatives in the early/mid 1980s. The business
reasoning, computational modelling, autonomous systems community became engaged as private companies
and robotics. The United States Defense Advanced Research boosted funding for AI research and development (R&D).
Projects Agency (DARPA), the National Research Council
However, the boom times of the 1980s were
and the United Kingdom Government were among the
followed by a second winter in 1987–1993. This was
more notable investors in AI capability. The 1960s can
triggered by the business community which
be considered AI’s first boom time. However, sentiment
increasingly felt their investments in AI were failing
changed in the early 1970s. The first AI ‘winter’ lasted
to achieve commercial outcomes. Similar to the
from 1974 to 1980. It was triggered by the Lighthill Report
first winter, there was again a prevailing sentiment
commissioned by the British Government and written
that AI had over‑promised and under-delivered.

3
Expectations had risen higher than what was achievable. In 2021 alone, 344,000 journal papers, books, book
It was reported that by 1993 over 300 AI companies had chapters and conference papers were published on the
shut down or gone bankrupt [23, 31]. This triggered a topic of AI [33]. In 2020 Google Scholar reported that AI
review of AI investment by governments; and again, AI attracted more citations than any other research field
R&D funds were suddenly and substantially reduced in the and five of the seven top-cited papers were on AI topics
United States, the United Kingdom and across the globe. [1]. Patents for AI have also been increasing sharply.
According to data from The Lens the number of published
Despite the two winters, the field of AI found a return to patents worldwide on the topic of ‘artificial intelligence’
growth in the 1990s seeing the rise of new paradigms, rose from 11,000 in 2017 to 57,000 in 2021 representing
tools, theories and applications. The 1990s saw the an average year-on-year growth of 84% over the last 5
rapid growth of the internet, data and computing years [32]. This growth is happening in all world regions
power. Since the late 1990s AI has remained on a strong and most countries, with China being a standout: the
growth trajectory. Research, investment, capability number of peer-reviewed publications on AI from China
and adoption have continued to expand; there hasn’t now exceeds that from both the United States and Europe.
been a third winter. There are no apparent signs of Furthermore, the growth in AI publishing is happening
a slowdown. In the current era AI is having a greater in all industry sectors with sharp increases in the
impact on scientific research than ever before. corporate, government, medical and other sectors [14].

In terms of expenditure, we are seeing sustained growth

2.2 Why is now different? which is likely to continue over coming years. Most
Many of the historic conditions which characterised governments from advanced economies have announced
the time periods leading up to previous AI winters, and funded significant AI strategies, roadmaps, plans
sometimes called AI ‘springs’, exist at the current and policies. Canada was among the first of the OECD
time. There has been a huge and sudden boom in countries to commit to a national AI strategy in 2017 [34].
investment. There is much hype. Expectations are Since then over 700 AI policy and strategy initiatives
running high, and there is considerable mythology and have been developed across 60 countries and territorial
confusion surrounding AI’s capabilities and functions. jurisdictions [3, 4]. By late 2019, over $86 billion in funding
If AI again fails to deliver on its perceived promises, it had been announced for AI initiatives [35]. Investment
may enter another winter. However, there are reasons has continued to grow. In 2021 worldwide spending on AI
to believe the current era is different from the past. In products and services grew 15.2% year-on-year, reaching
this section we briefly explore what is different about US$341.8 billion. Growth of 18.8% is expected for 2022
the current epoch in the timeline of AI for science. with total spending forecast to exceed US$500 billion
per year by 2024 [36]. The share of this expenditure being
invested in science and research is unknown, but due to
2.2.1 Greater depth and breadth the novelty and complexity of developing AI – e.g. training
of adoption machine-learning algorithms – it is likely to be substantial.

Compared to historic booms, today’s AI surge has It is also worth noting that AI has now found its way
greater depth and breadth of technology penetration into people’s day-to-day lives. Billions of people use,
within diverse scientific fields, industry sectors, and increasingly depend upon, AI on a regular basis.
geographies, policy spheres and demographics. AI has Countless companies use AI technologies to provide goods
gotten into practically everything everywhere. This is and services to their customers. Before the turn of the
creating greater resilience for AI compared to historic century this was not the case. In the 1990s (and before)
boom–bust cycles. AI is too deeply embedded in too AI was a concept beyond the realm of most people’s
many places to suddenly lose relevance, as happened lived experience. In comparison, people today routinely
in the two winters of 1974–1980 and 1987–1993. interact with powerful AI through smart-phones, smart-
cars and smart-speakers. This makes AI both tangible
This embeddedness can be seen through publishing and
and practically useful. The contemporary widespread
patent trends. Our bibliometric analysis based on data
familiarity with AI makes it easier for today’s research
from The Lens [32] reveals that 5.7% of all peer-reviewed
community to communicate its value proposition.
research publications refer to AI in the title, abstract or
keywords. This is up from 3.1% in 2017 and 1.2% in 2000.

4 Artificial intelligence for science

2.2.2 Hardware, software and 2.2.3 Commercial drivers
data availability While there have been surges in AI investment in the past,
Throughout history AI scientists have struggled they do not come close to what is happening today. Private
to implement and test their ideas due to a lack of investment flows into AI have increased substantially over
computational resources and/or lack of data to train the past several years (Figure 2). Despite the pandemic,
machine-learning algorithms. Much was theorised but private investment in AI companies increased by a record
could not be proven nor developed. That is often why high of 9.3% in 2020 year-on-year – which is above the 5.7%
scientists hit a wall and could not realise their objectives. increase of 2019 – and exceeded US$40 billion [14]. Venture
This boom is different. This boom comes with much capital investment in AI has also been growing compared
better tools and much better data. Scientists can turn to other areas of investment. According to the OECD, the
their ideas into technologies and technologies can be share of venture capital investments in AI start-ups reached
turned into consumer products with commercial value. 20% of all venture capital investments in 2020, up from
3% in 2012 [37]. The number of venture capital deals in AI
For example, in the last 10 years we have seen the rise companies grew by 34% annually between 2012 and 2020
of graphics processing units (GPUs) which are well from 500 deals in 2012 to 3,900 deals in 2019 [37]. In 2020
suited to support parallel computing. The GPU has Australia was ranked 11th in the world by the total amount
been transformative for AI. It has enabled low-cost of private investment in AI companies [14] with the United
high-power computing for a vast range of complex States, China, United Kingdom and Israel at the top of the
machine-learning challenges. Furthermore, cloud- list. With so much invested, AI activity by R&D providers and
based computing services are bringing these within product developers is likely to be sustained for some time.
reach. We are also seeing software tools and platforms,
such as PyTorch, Tensorflow, Theano, MxNet, Microsoft
Azure, and Amazon Web Services, make AI much more
accessible to a broader cross‑section of scientists; a 42.24

subset of AI functionality is usable for those without 38.66

36.58
highly developed and specialised AI skills. More recently
we have seen code-free AI tools that allow users to
perform AI functions via relatively simple graphic user
interfaces (GUIs). These are likely to improve over time.
23.00
Another limiting factor for AI scientists in history has been
the availability of data to train machine-learning algorithms.
However, there’s no shortage of data today. There are
13.10
challenges about managing an overwhelming volume,
variety and velocity of data. There are also challenges 7.95
about verifying data, as well as challenges about handling
private and confidential data. If these challenges can be
addressed, today’s scientists can have access to more
data on every topic than ever before in history. The data 2015 2016 2017 2018 2019 2020
comes from human internet usage, sensory systems and
Figure 2. Private investment in artificial intelligence companies
countless other rapidly expanding sources. These data worldwide (billions of US dollars).
are providing scientists with new opportunities to use AI Data source: Stanford University Artificial Intelligence Index [14].
to identify patterns, test hypotheses and make predictions.

5
2.2.4 Improved scientific knowledge and backpropagation algorithms which improve the accuracy of
technological capability artificial neural networks by finely adjusting mathematical
weight functions. The recent transition to the Rectified
Lastly, today’s surge in artificial intelligence comes with Linear Unit (ReLU) activation function has substantially
solutions and/or improvements to some – but by no means helped address the vanishing gradient problem; a
all – of the longstanding machine-learning problems, and longstanding challenge in the field of machine learning [14].
with knowledge gaps filled which limited AI’s development
historically. For example, CSIRO re-implemented a classic Solutions to these, and other, AI barriers have opened
machine-learning algorithm, Random Forest, enabling up entirely new avenues for continued problem solving
it to overcome the ‘curse of dimensionality’, which and improvement of AI technologies. This means
was brought on by today’s larger and more detailed AI science has a greater chance of delivering on
datasets [6]. Other algorithmic improvements include expectations. The future is likely to see continued
new regularisation techniques which modify the learning discovery and innovation in the field of AI enabling the
algorithm to improve the generalisability of the model development of enhanced technological capabilities.
plus improve its performance on unseen data [38].
These techniques can reduce the problem of ‘overfitting’
2.2.5 No slowdown in sight
in machine learning which happens when the model is
too closely matched to input data and is, therefore, unable There’s so much momentum behind the current AI
to predict future observations accurately. We have also growth cycle it is hard to see it ending anytime soon.
seen the emergence of robust optimisers such as Adam If AI were to experience another winter in one field of
[39], RMSprop [40] and modification of the stochastic research – such as computer science – it is unlikely to
gradient descent (SDG) procedure [41]. These approaches be winter everywhere. The field of AI has become so
speed up optimisation algorithms and generate higher large and diverse it is likely to be experiencing all four
quality solutions compared to earlier methods. The last seasons in sub-fields, application domains, geographies
10 years have also seen the emergence of improved and industry sectors at any one point in time.

HairNet is an artificial intelligence and

machine learning model that can score leaf
hairiness in cotton to assist breeders in
identifying plants with beneficial traits.

6 Artificial intelligence for science

3 The science sector
In this section we describe the size, structure and trends within the Australian
and global science sectors. We examine how the sector has expanded and we
examine the science productivity slump. It is likely that AI will play an important
role in boosting science productivity and, in turn, economy-wide productivity.

3.1 What is the science sector? 3.2 The global science sector
The science sector is an interconnected, collaborative and For some time, the world has been growing its science
dynamic global community with highly porous boundaries. workforce and research spending (Figure 3). In 2018 the
It captures a diverse range of conceptual frameworks, global research workforce was estimated by the United
paradigms, methodologies and cultural approaches to Nations at 9.33 million workers up from 8.01 million in 2014.
knowledge discovery. The Australian Academy of Science This workforce is estimated to be growing over three times
[42] says, ‘science can be thought of as both a body of faster than population growth; increasing by 16% during the
knowledge (the things we have already discovered), and the 5-year period 2014 to 2018 [44]. Data from the OECD show
process of acquiring new knowledge (through observation that the number of researchers per 1,000 employed persons
and experimentation—testing and hypothesising)’. has increased from 6.1 to 8.9 during 2000 to 2020 [2].

Science happens in households, communities, Expenditure patterns for R&D are indicative of aggregate
start‑ups, large companies, government agencies, global-level science spending. Spending on R&D is
research institutions and universities. There are many outpacing global economic growth, reaching 2.2%
ways of classifying scientific activity. The Australian of global GDP by 2020 compared to 2.0% a decade
Academy of Sciences identifies four broad and ago. In the OECD it has grown faster, reaching 2.48%
high‑level categories of scientific research [42]: in 2019. Recent estimates suggest the world now
spends over US$1.7 trillion per year on research with
• natural science – the study of living organisms 10 countries accounting for 80% of expenditure. The
and physical sciences which includes the top five spenders in absolute terms are the United
study of the material universe States, China, Japan, Germany and South Korea. The
• social science – the study of human individuals, country spending the most on R&D relative to GDP is
communities, societies and institutions Israel, which invested 4.93% of GDP in R&D in 2019 [2].
and how they interact and behave
Global technology corporations are another key source
• formal science – the study of logic and mathematics
of expanding funds, and overall activity, for scientific
• applied science – the adaptation and research. This has been a relatively recent phenomenon
use of existing scientific knowledge for and is especially relevant to AI science. During 2005–2018,
industry and societal applications. global private-sector R&D spending rose from $523.8 billion
The United Nations recently conducted global consultation to $1.1 trillion [45]. A NASDAQ report [46] reveals the R&D
to define the science sector for statistical purposes [43]. budgets of the top spenders in 2020: Amazon ($62.3 billion),
This analysis identified the concept of ‘scientific and Alphabet ($40.2 billion), Huawei ($32.1 billion), Microsoft
technological activities’ which includes three components ($28.1 billion), and Apple ($27.3 billion). These trends suggest
of science activity: (a) research and experimental that private corporations are funding or doing much of the
development, (b) scientific and technological education world’s scientific research. Given the business objectives
and training, and (c) scientific and technical services. of these companies, much of this R&D is likely within the
We have defined and conceptualised the science sector fields of data science and AI. The entry of the private/
in the same manner in the subsequent analysis. corporate sector into R&D is impacting scientific research.

7
EXPENDITURE ON R&D (US DOLLARS, TRILLIONS) PERCENT OF GDP SPENT ON R&D

1.6 2.6

USD Trillions (current prices)

1.4 2.5
Percent of GDP

1.2 2.4

1.0 2.3

0.8 2.2

0.6 2.1

0.4 2.0

0.2 1.9

0 1.8
2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019
Figure 3. Increasing spending on R&D within the OECD.
Data source: OECD Statistics [2].

The bulk of the global research effort by publishing The volume of research published across all fields of
volume lies in the physical sciences, which contained science has been increasing consistently for over 60 years
44% of all publishing in the year 2021. This is followed by (Figure 5). However, during the peak COVID-19 pandemic
the health sciences, life sciences and the social sciences years of 2020 to 2022, global combined scientific publishing
and humanities, which contained 22%, 20% and 14%, via journal papers, conference papers, books, book
respectively, of publishing in the same year. The relative chapters and dissertations declined by 12%. This may have
level of publishing within these domains has remained multiple causes, including: (a) the redirection of research
relatively stable over history. When we look beneath effort towards urgent COVID-19 issues; (b) a switch to
the top-level research domains, we see that the fields alternative non-traditional publishing venues; (c) decreased
of medicine, biochemistry genetics and molecular productivity of researchers working in lockdowns; and (d)
biology, engineering, social sciences and computer reduced research funds in the university sector associated
science account for over half of all research publishing with reduced revenue due to COVID-19 disruptions.
(Figure 4). Medicine is by far the largest research field, The patterns and consequences of pandemic-related
accounting for one-fifth of all research publishing. research activity contraction are explored in the research
Again, this has remained relatively stable over history. literature in greater depth [48, 49]. These studies indicate
the possibility of long-term effects and the disproportionate
impact on female researchers with young children.

8 Artificial intelligence for science

Agricultural and biological sciences 3.6%
Arts and humanities 2.2%
Biochemistry genetics and molecular biology 10.1%

Business management and accounting 1.7%

Chemical engineering 2.4%

Chemistry 5.4%

Computer science 5.7%

Decision sciences 0.5%

Dentistry 0.2%

Earth and planetary sciences 2.7%

Economics and finance 1.1%

Energy 2.6%

Engineering 11.2%

Environmental science 5.3%

Health professions 0.5%

Immunology and microbiology 2.1%

Materials science 5.5%

Mathematics 3.3%

Medicine 17.8%

Neuroscience 1.5%

Nursing 1.3%

Pharmacology toxicology and pharmaceutics 2.4%

Physics and astronomy 3.8%

Psychology 1.3%

Social sciences 5.5%

Veterinary 0.3%

Figure 4. Share of global peer-reviewed publishing by research field in 2022.

Data source: The Lens [32] and Scopus, Elsevier All Science Journal Classification [47]. Data sourced for 1 Jan 2022 to 20 Sept 2022.

9
NUMBER OF PUBLICATIONS (THOUSANDS)
8,000

7,000

6,000

5,000

4,000

3,000

2,000

1,000

0
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Book Book chapter Conference paper Journal article

Figure 5. Worldwide peer-reviewed scientific publishing in all fields for 1960–2021.

Data source: The Lens [32] database of peer-reviewed scholarly publications.

Publishing venues may also be changing. During recent 3.3 The Australian science sector
times, new open-access research-sharing platforms have
risen quickly. For example, during 2016–2021 submissions In the Australian and New Zealand Industry Classification
of research papers to the arXiv pre-print server increased (ANZIC) the science sector fits under the industry
60% from 113,380 to 181,630 [50]. The material is not subject sub‑grouping ‘scientific research services’ [53]. In 2016, this
to peer review and, therefore, comes with a significant industry contained 28,850 workers, roughly 0.3% of the
quality and reliability caveat. However, it does allow Australian workforce. However, most scientists work within,
for rapid sharing of research results and is being used and most science happens within, other industry sectors
extensively. In another example, the ‘Papers with Code’ (Figure 6). The vast bulk of Australian scientists do not work
platform now reports 62,856 research papers (as of 8 Jan in the science industry. They use science to problem‑solve
2022) [51]. The extent to which these alternative venues in other industries and societal spheres. The ‘science
displace and/or complement traditional peer-reviewed sector’ is spread widely across the entire economy.
publishing, and/or morph into new publishing models,
Using the most recent population census for 2016,
remains to be seen. Peer review is still needed to separate
we identify 109,890 natural, physical and social scientists
knowledge from opinions. The question is whether new
in Australia representing 1% of the total workforce [53].
models can uphold this critical function whilst speeding
We note a degree of uncertainty in this estimate due to
up the process of knowledge sharing [52]. What’s evident
the category groupings used by the Australian Bureau
at the current time is sizeable and rapidly increasing usage
of Statistics. It is likely that we are including some
of the alternative venues by the science community.
non‑scientists and excluding some scientists because
there is no category for ‘scientist’ in our statistics; it
requires aggregating and disaggregating other categories.

10 Artificial intelligence for science

Nevertheless, we believe a reasonable approximation as management consulting, architectural, engineering,
has been achieved using the occupation information law, accounting, market research, veterinary and other
in the census (Appendix A). From this we identify 11 services. This analysis suggests that much of AI for science
sub‑categories of the scientist profession (Figure 7). (the focus of this report) is likely to occur within Australian
The two largest groupings are medical laboratory industry as opposed to dedicated research organisations
scientists and environmental scientists, which together and universities. This is because most scientists work in
account for about 30% of the scientific workforce [53]. industry. However, research organisations may be more
focused on theoretical and early-stage developmental
At 22% of the total science workforce, Australia’s aspects of AI compared to industry scientists who
professional services sector contains the largest share may be focused on adoption and application.
of scientists [53]. It includes sub-industry groupings such

Agriculture, forestry and fishing 3,460

Mining 6,157

Manufacturing 7,551

Scientific research services 9,263

Public administration and safety 12,902

Education and training 13,894

Health care and social assistance 15,793

Other industries (12 categories) 16,541

Professional and technical services 24,329

Figure 6. The number of Australian scientists by industry grouping.

Data source: Australian Bureau of Statistics, 2016 Population Census [53].

Life scientists 4,966

Agricultural and forestry scientists 5,438

Mathematicians and statisticians (includes actuaries) 5,604

Chemists, and food and wine scientists 6,605

Geology, geophysics and hydrogeology scientists 6,811

Veterinary scientists 8,499

Social scientists (geography, economics, anthropology,... 12,059

Science technicians 13,025

Unclassified natural and physical science professionals 14,557

Environmental scientists 16,041

Medical laboratory scientists 16,277

Figure 7. The number of Australian scientists by field of science.

Data source: Australian Bureau of Statistics 2016 Census (see Appendix A for methods and assumptions) [53].
Note: Thesocial science field contained a small number of non-science occupations that could not be disaggregated.

11
R&D spending patterns recorded by the Australian Bureau A specialisation quotient above 1 implies the research field
of Statistics [54] provide insight into science investment is associated with greater output in Australia compared to
patterns. In 2019–2020, total expenditure on R&D the average for all research fields. This analysis reveals that
by business, government, higher education and not- over the 10-year period 2012–2021, the fields of psychology,
for‑profit sectors in Australia amounted to $35.602 billion, health professions, earth and planetary science, nursing,
representing an increase of $2.54 billion (8%) over the agricultural and biological science and environmental
preceding 2 years. Much of the growth in R&D spending science have higher levels of comparative specialisation
comes from the higher education sector, which increased in Australia (Figure 8). By comparison, Australia has lower
from $11.24 billion in 2017–2018 to $12.71 in 2019–2020, levels of comparative specialisation in dentistry, physics and
an increase of $1.47 billion (13%). However, over the astronomy, mathematics, chemistry and materials science.
longer-term Australian R&D spending declined as a At the even more granular (third) level of the research field
percentage of GDP from 2.1% in 2012 to 1.8% in 2020. hierarchy, the top 10 fields of research by specialisation for
Australia include: emergency medical services; research and
At the higher levels of the research field taxonomy, the theory; community and home care; tourism, leisure and
publishing profile for Australian research has similar hospitality management; economic geology; chiropractics;
expression to global research. However, at a more occupational therapy; pharmacy; physical therapy, sports
granular level we can see that Australia has comparative therapy and rehabilitation; and ecological modelling.
specialisation in certain fields. To examine this, we
calculate specialisation quotients at the second-level of
the All Science Journal Classification (ASJC) by Scopus,
Elsevier [47] which contains 26 unique research fields.
The specialisation quotient is calculated as follows:

Number of publications Number of publications

in research field by ÷ in all research fields by
Australian authors Australian authors
Specialisation quotient
for a given research =
field in Australia
Number of publications Number of publications
in research field by ÷ in all research fields by
worldwide authors worldwide authors

12 Artificial intelligence for science

Agricultural and biological sciences 1.32

Arts and humanities 0.75

Biochemistry genetics and molecular biology 1.04

Business management and accounting 1.27

Chemical engineering 0.71

Chemistry 0.70

Computer science 0.83

Decision sciences 1.01

Dentistry 0.58

Earth and planetary sciences 1.42

Economics econometrics and finance 1.21

Energy 0.92

Engineering 0.81

Environmental science 1.31

Health professions 1.54

Immunology and microbiology 1.02

Materials science 0.72

Mathematics 0.69

Medicine 1.06

Neuroscience 1.29

Nursing 1.37

Pharmacology toxicology and pharmaceutics 0.78

Physics and astronomy 0.64

Psychology 1.75

Social sciences 1.21

Veterinary 0.98

0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0

SPECIALISATION QUOTIENT

Figure 8. Comparative levels of specialisation in Australian science and research during 2012–2021.
Data source: The Lens [32] and Scopus, Elsevier All Science Journal Classification [47]. A higher score indicates greater
specialisation within the given field of research compared to the average comparison of Australia and the world.

13
3.4 Productivity decline A comprehensive study published in the American
Economic Review in 2020 finds that within the science
Productivity is the efficiency via which inputs are sector ‘research effort is rising substantially while research
converted to outputs. Productivity is an important productivity is declining sharply’ [56]. The analysis
determinant of short-run economic growth and the primary examines agricultural crop yields, semi-conductors, cancer
determinant of long-run economic growth (and wealth treatments, heart disease treatments, intellectual property
generation). Ideas are the fuel source for productivity. patents and overall economic productivity. The analysis
When people discover how to produce something or uses outcome metrics related to benefits such as changes
deliver a service more efficiently it leads to productivity in crop yields (e.g. wheat, corn) per unit of area resulting
growth. Science is one of the most important pathways to from agricultural R&D. Similar outcome metrics were
discovery and ideation. The positive associations between used for the other categories. The researchers show that
science, ideas and productivity are well established the cost of developing new pharmaceutical products to
and accepted, although the magnitude of benefit is not treat illness doubles every 9 years. They find that while
easily quantified [55]. However, over the past decade or the research workforce has grown, productivity (output
so, the world’s advanced economies, and the science per researcher) has decreased [56]. We are getting
sector itself, have been caught within a productivity fewer ground-breaking ideas for each dollar invested.
slump (Figure 9). This is harming economic growth and For example, the cost of developing a new antibiotic
limiting long-term improvement of living standards. was estimated at US$1.581 billion in 2017 [57]. This far
exceeds the costs of antibiotic discovery compared to
the ‘golden era’ of the 1970s and 1980s. For antibiotic
1.4% discovery, and many other types of scientific discovery, the
1.2% 1.2%
next wave of discovery appears harder to achieve [58].
1.0%
The authors conclude that the United States needs to
double research effort, and double its research workforce,
0.4% 0.5% every 13 years to maintain science output. Without this
investment, they argue, the United States economy will
experience productivity decline and declining rates of
GDP growth [56]. The United States economy, and other
advanced economies worldwide, increasingly depend
-0.2% on science and technology improvements to sustain
growth. Therefore, it is critical that the science sector
1985–89

1990–94

1995–99

2000–04

2005–09

2010–14

2015–19

– the engine room for the creation of ideas – keeps

operating at full pace. If science productivity is declining,
the only way to achieve this is via investing in more
Figure 9. Average (mean) annual OECD-country multi-factor scientists and more science resources. The economists
productivity growth.
who did this study indicate that in advanced economies,
Data source: OECD Statistics [2].
income growth and improved living standards depend
on research productivity and research effort [56]:

Economic growth = Research productivity × Number of researchers

e.g. 2% to 5% per year decreasing increasing

14 Artificial intelligence for science

The centrality of innovation, science and research to There will be many policy interventions needed to solve
productivity uplift and economic growth is well accepted the productivity slump in science. However, the recent
and demonstrated within the field of economics. It means surge in AI capability and adoption is likely to play an
that if research productivity is declining, the only way important role. Recent years have seen AI substantially
of ensuring income growth is via increasing research improve the speed, quality, safety and cost-effectiveness
effort. This is the approach followed by most advanced of scientific research. AI is already enabling discoveries
economies, as shown in the R&D expenditure data and which were hitherto beyond reach. Although AI has
statistics on the growing R&D workforce presented earlier been used by scientists since the 1960s, it hasn’t been
in this chapter. Most OECD economies are growing the mainstreamed until the last several years. The last few
share of GDP spent on R&D to offset declines in research years have seen a huge increase in AI development
productivity and achieve overall economic growth. and application in all scientific fields. AI is likely to be
one of the most important mechanisms for boosting
The United States economic study finding productivity science productivity and escaping the slump. The need
decline isn’t a standalone. A study from the Research for science to reinvent itself, and problem-solve for
Institute of Economy, Trade and Industry in Japan applied industry and society, is likely to be a driving consideration
the same techniques and found ‘significant decline of R&D for AI development and adoption into the future.
efficiency in the Japanese information service industry’ [12].
The researchers recommend that the Japanese government
implement R&D policies that address the decline. Another
team of economists from the Leibniz Centre for European
Economic Research and the Copenhagen Business School
replicated the United States study for China and Germany
using firm-level data over three decades [13]. They find
evidence of productivity decline in both countries and
‘strong decline’ in Chinese R&D productivity. The authors
conclude that ‘diminishing returns in idea production are
a global phenomenon, not just confined to the United
States’ [13]. These studies were presented at a recent OECD
workshop examining science productivity and AI [11].

15
Our bushfire simulation
software Spark models bushfire
spread to help plan for and
manage bushfires.

16 Artificial intelligence for science

4 Artificial intelligence and
knowledge discovery
4.1 Enablement or transformation? This is a well-understood but extremely difficult‑to‑solve
optimisation problem for nuclear physicists. Reinforcement
To demonstrate the potential of AI to impact knowledge learning was able to identify plasma configurations
discovery, let us consider the case of electricity generation not previously known. As reported in Wired magazine
via nuclear fusion. According to the International [62], Ambrogio Fasoli (fusion and plasma physicist
Thermonuclear Experimental Reactor organisation in and director of the Swiss Plasma Center) says this
France, nuclear fusion creates 4 million times more represents a ‘significant step’ on the pathway to nuclear
energy than chemical reactions such as burning coal, fusion and that AI enables ‘us to explore things that
oil or gas [59]. Electricity generation via nuclear fusion we wouldn’t explore otherwise, because we can take
represents one of the most important (future) scientific risks with this kind of control system we wouldn’t dare
discoveries for humanity. If discovered, nuclear fusion take otherwise’ and that ‘if we are sure that we have a
would provide an abundant and practically inexhaustible control system that can take us close to the limit but not
source of clean energy. Nuclear fusion does not produce beyond the limit, we can actually explore possibilities
the high-activity and long-lived nuclear waste associated that wouldn’t otherwise be there for exploring’ [62].
with nuclear fission. Nuclear fusion provides greatly
enhanced safety, substantially reduced financial costs For scientists working on nuclear fusion, AI has provided
and reduced risk of weaponisation [59]. Nuclear fusion a big boost. It has removed one of the critical barriers on
can help solve climate change while supplying abundant the pathway to discovery; the ability to control the plasma
energy for ample food, water and mineral (via mining and within the tokamak. The future impact of AI on science
recycling) production. It is a game changer for humanity. and knowledge discovery can be viewed as a continuum
of possibility. At one end of the continuum is enablement:
The main problem with electricity generation via nuclear the useful application of AI tools to help scientists do what
fusion is that it currently cannot be accomplished in a they are already doing faster, cheaper, safer and better.
practical and industrial way. The scientific community has At the other end of continuum is transformation: the use of
been trying for decades. However, it appears this capability AI to remove major barriers to scientific progress leading
is getting closer. In early 2022 British scientists reported to paradigmatic shifts, new approaches to knowledge
the production of 59 megajoules of energy sustained for discovery and new possibilities for problem solving.
5 seconds from a nuclear fusion reaction; while the duration
is minuscule by industrial electricity generation standards, The case for enablement is well demonstrated through
this is nevertheless a major improvement upon previous thousands of published AI studies within practically all
records [60]. Another significant breakthrough came from science fields over recent decades. Our observations in
the field of AI at about the same time. In February 2022 this report, that the share of global scholarly publishing
the results of a collaboration between DeepMind and on AI has risen since 2020 from 1.2% to 5.7% and AI
the Swiss Plasma Center were published in Nature [61]. is now applied in virtually all disciplines, provide
In this project reinforcement learning, a type of machine evidence of the usefulness of AI to scientists and
learning, was used to control the super-heated suspended researchers. The extent to which AI will (in the future) be
plasma needed for the nuclear fusion reaction within a transformative and associated with paradigmatic shifts
device called a tokamak. Reinforcement learning was used in approaches to knowledge discovery and major leaps
to control voltage in the tokamak and, thereby, the shape in problem‑solving capability is less clear. However, some
of the suspended plasma, ensuring it met experimental AI scientists see this as a distinct possibility.
requirements while not touching the walls of the tokamak.

17
Hiroaki Kitano is a Japanese AI scientist and is the director data-intensive approaches is they (a) fail to draw upon
of both the Systems Biology Institute and Sony Computer existing theory, and (b) fail to establish and understand
Science Laboratories in Tokyo. Writing in Nature [63], causality [73]. This can lead to errors and accidents.
Kitano proposes the ‘Nobel Turing Challenge’, which ‘aims
to develop a highly autonomous AI system that can perform Recent perspectives suggest that ‘data versus theory’
top-level science, indistinguishable from the quality of is a false dichotomy and that there is no competition
that performed by the best human scientists, where some between the two [69]. Instead, data science approaches
of the discoveries may be worthy of Nobel Prize level are inextricably linked to theory and have unique
recognition and beyond’. In this paper Kitano also argues application in practically every field of science [74].
that future AI science ‘may be an alternative form of science Techniques such as linear regression analysis have long
that will break the limitation of current scientific practice been used by scientists to understand and explain
largely hampered by human cognitive limitation and real‑world phenomena. When linked with other tools
sociological constraints’ and that such approaches ‘could and ideas, these techniques can lead to confirmation of
give rise to a human-AI hybrid form of science that shall existing theories or the development of new theories.
bring systems biology and other sciences into the next Scientists have always used data, and as the tools of
stage’ [63]. Kitano is connected with a team at the Alan data science get better, they are getting better at using
Turing Institute in the United Kingdom which is delivering data. Theory and causality haven’t vanished; they remain
a project (‘The Turing AI scientist grand challenge’) critically important. The observations about data-intensive
which tackles similar and related objectives [64]. As stated science are likely to apply to AI-intensive science. Data
on the Alan Turing website, this ongoing project started science often uses AI, and AI almost always uses data
in January 2021 and is, among other things [64, 65]: science; the two fields are increasingly inseparable.

• reviewing current autonomous systems capable

of performing scientific research with a focus on 4.2 Case studies – Artificial
AI approaches capable of pushing disciplinary intelligence applications for science
science beyond the current cutting-edge
In this section we describe case studies where
• developing a multi-year roadmap charting AI has improved the efficiency and effectiveness
a scientific and technical pathway for AI for (the productivity) of scientific research. A more
science with milestones identified in materials, detailed, comprehensive and up-to-date repository
biomedicine and environmental sciences. of case studies is available on the CSIRO website
(www.csiro.au/en/research/technology-space/ai).
A related line of inquiry pursued by researchers working
on the philosophy of science is about the possibility The case studies also illustrate how AI is enabling science
of theory-free, data-intensive science [66-68]. Starting and research in diverse fields of study. This includes
in the early 2000s ‘this approach is supposed to be examples of enablement and transformation, where
data driven, strongly inductive, and relatively theory AI has enabled scientists to solve complex problems
independent’ [69]. Data-intensive science can be and has created an elevated platform of capability and
considered transformative as it represents a paradigm knowledge discovery. Overall, AI and its constituent
shift challenging existing approaches to knowledge technologies (such as pattern recognition and machine
discovery. The idea comes from the successful application learning) appear to considerably enhance and accelerate
of data science for forecasting in fields like meteorology the scientific process by allowing: (a) faster processing of
[70], economics, energy, and demographics [71]. data, (b) handling of very large and datasets, (c) handling
The idea has credence because sometimes data can of disparate datasets, (d) offloading of menial tasks,
work better than theory for modelling and predicting (e) deeper and wider exploration of the experimental
system behaviour [71]. However, science philosophers space, (f) more accurate predictions due to better models,
have criticised the idea of theory-free data‑intensive and (g) faster and more reliable detection of salient
science [72]. Some have argued that researchers using and/or anomalous patterns or events. However, the
‘big data’ approaches (akin to data-intensive science) enhancement and acceleration of science via AI is largely
may sidestep the critical hurdle of causality and rely predicated on relevant data being available in digital
upon statistical correlation to explain and predict system format, thereby necessitating that any physical experiments
behaviour [73]. The concern about sole reliance on need to be designed and executed such that the primary
objectives include data acquisition in digital format.

18 Artificial intelligence for science

4.2.1 Predicting the 3D structure 4.2.3 Enhancing the reach of citizen science
of proteins AI is being applied in citizen science, assisting civic
Proteins are essential for the growth and maintenance of educators and scientists in engaging the community in
all cells and tissues [75, 76]. Understanding their structure, scientific endeavours and collecting large datasets on
or the way they fold, is key to identifying their function – rare or difficult to access phenomena. Examples include
a time-consuming and challenging problem that scientists iNaturalist, a platform run by the California Academy of
have spent decades trying to solve [75, 76]. AlphaFold Sciences and National Geographic where members of the
is a neural network system developed by Google’s public can submit photos of the natural world, including
DeepMind that can map the 3D structure of proteins animals and plants [79]. This platform uses computer
with significantly greater accuracy than conventional vision and a machine-learning model previously trained
methods [75, 77]. AlphaFold has been applied to map on an existing research-grade dataset of images [80, 81].
the human proteome (the entire set of proteins that Citizen science can enhance the spatial and temporal
make up the human body) and was able to predict the resolution of data in ecological monitoring projects
structure of 98.5% of human proteins [77]. Together with relative to traditional methods [82]. Using AI systems can
the European Molecular Biology Laboratory’s European improve the cost efficiency of collecting, processing and
Bioinformatics Institute, DeepMind has developed the analysing data generated by the public [82] enabling more
AlphaFold Protein Structure Database to make these researchers to leverage the benefits of citizen science.
predictions available to the scientific community [78].
This technology can potentially support future advances
4.2.4 Predicting the replicability of
in biological research and drug development.
scientific studies
The replicability of scientific findings was brought into
4.2.2 Accelerating solar panel research
question with a series of publications demonstrating that
Researchers at CSIRO have developed a research robot a large share of studies in psychology, economics, and
that can autonomously test flexible solar panel samples medicine could not be replicated [83-85]. Non-replicability
[5]. These researchers developed the autonomous system can impede scientific progress, hinder public support and
during the ‘second wave’ of COVID-19 in Melbourne in trust in science, and waste finite funding resources [86].
2020. Before this, researchers could manually test up to Researchers from Northwestern University used machine
20 solar cells per day and had to be physically present learning to accurately estimate the replicability of a study;
in the lab [5]. The new automated research system is meaning the extent to which it is possible to replicate
controlled remotely and could test 12,000 cells in 24 hours, the methodology but not necessarily whether (or not)
which represents a 600-times improvement in productivity the results hold-up. The machine-learning approach
[5]. AI and machine learning is also being applied to performed as well as expert survey predictions, which
efficiently analyse and predict parameters for solar cell is the current gold-standard, but resource-intensive,
manufacturing of organic solar cells [77]. These applications method of assessing replicability [86]. While this research
illustrate how autonomous testing, combined with machine is preliminary, it suggests that AI can potentially be used to
learning, can increase the efficiency of scientific research test the replicability of scientific findings without imposing
and accelerate the development of new technologies, additional time and resource requirements on scientists.
even while scientists are working from home [5].

19
4.2.5 Discovering and developing To provide better intelligence, deep-learning techniques
new materials have been used to count the number of endangered species
animals from aerial survey images [92, 93]. Using images
AI is being widely applied across materials science to collected from motion-sensor cameras placed in natural
accelerate the rate at which scientists can discover and habitats, researchers found that a deep-learning system
develop new materials [87]. An early example is the can identify animals as accurately as a human observer
Autonomous Research System developed by researchers [93]. The use of AI here to classify around 5.5 million
from the Air Force Research Laboratory, UES Inc. and images saved over 8.4 years in human labour [93].
Lockheed Martin Advanced Technology Laboratories This demonstrates the significant cost and time savings
[88]. This system combines robotics, AI, data science, that AI can provide in conducting conservation research.
and in situ technologies to design, execute, and analyse
experiments faster than traditional human-driven
4.2.8 Predicting high-impact research
approaches [88]. This approach has been used to explore
the synthesis of carbon nanotubes – a well-suited material The impact of scientific outputs is typically measured
for electronics applications that scientists have spent through citation metrics, such as h-indices and journal
decades trying to understand [88]. Through autonomous impact factors. These metrics can be discipline‑specific,
experimentation, AI can increase the speed and cost of biased, or reflect lag quality indicators [94, 95].
materials science research, increase the productivity of With trillions of dollars invested in research globally each
scientists, and maximise the value that can be derived year, having reliable predictors of impactful research is
from complex multi-dimensional datasets [87]. critical. Researchers from the Massachusetts Institute
of Technology have developed a DELPHI framework
4.2.6 Untangling mathematical (Dynamic Early-warning by Learning to Predict High
Impact), which uses machine learning to predict the
relationships
likely impact of scientific publications [96]. It draws upon
Advances in mathematics depend on the ability of a rich collection of publication, journal and citation
mathematicians to discover new patterns and formulate data [96]. The model was able to predict high-impact
statements around the potential relationship between research the year it was published with 77% accuracy,
objects (referred to as a conjecture) [89]. These insights and it was a better predictor than citation metrics [96].
are then used to develop new mathematical proofs [89]. This work demonstrates how AI can potentially be used
AI can assist mathematicians in the initial step of detecting to inform funding decisions to maximise the return
patterns between objects, which can help guide them in on investment and impact of scientific research.
developing mathematical formulae and theorems [89].
A team of researchers from DeepMind, the United Kingdom 4.2.9 Decoding the human brain
and Australia have applied AI to study the algebraic
and geometric structure of knots – a longstanding The human brain is arguably the most complex system
mathematical challenge [90]. The machine-learning known to humankind, but AI is helping scientists to crack
approach enabled the researchers to discover novel and the neural code. AI provides scientists with opportunities
surprising patterns, and develop new conjectures [89]. to directly explore the functioning of healthy, neurotypical
human brains, which has previously been limited, if not
impossible, due to practical or ethical considerations
4.2.7 Improving the efficiency in
[97]. For example, researchers from the Massachusetts
conservation science Institute of Technology have used deep neural networks
Like many aspects of science, conservation research to demonstrate the hierarchical structure of the human
operates in a resource-constrained environment. visual system [98]. The neural-network model can
Conservation managers need to determine the most identify objects as well as a human and exhibits a similar
effective way of managing these finite resources and pattern of neural activity to a monkey brain performing
identify when to stop efforts to manage and survey an a similar task [98]. Similar applications have been used
endangered species population. Still, they often have to show the hierarchical organisation of the human
insufficient information for making these decisions [91]. auditory cortex [99]. These neural-process models can
help scientists generate hypotheses around certain brain
functions and inform their experimental design [97].

20 Artificial intelligence for science

4.2.10 Identifying drugs for Review Assistant, launched by open-access publisher
antibiotic‑resistant bacteria Frontiers in June 2020, which can screen the language
quality, integrity of figures, instances of plagiarism
AI has been applied in a range of medical contexts, and and potential conflicts of interest in submissions [108].
increasingly, it is being used to accelerate the discovery Other AI-enabled systems use natural language processing
of new drugs [100]. Improving the efficiency of drug tools to synthesise academic papers into language that
discovery research is particularly critical in the face of a 7-year-old child can understand [109]. These tools
the increasing prevalence of antibiotic-resistant bacteria can potentially improve the productivity of academics
and the diminishing returns on investment and high risk and publishers and assist them in identifying and
associated with drug discovery [101-103]. Researchers have evaluating relevant and impactful research [107, 108].
used deep learning to develop a model that can identify
candidate molecules capable of inhibiting the growth of
4.2.13 Identifying archaeological samples
Escherichia coli (E. coli) – a common bacteria associated
with antimicrobial resistance [101]. Similarly, scientists have In archaeology, deep-learning approaches have been
also used AI to rapidly identify and test novel compounds increasingly applied to make sense of often unstructured
that can inhibit discoidin domain receptor 1 (a common and disparate datasets, and to derive new insights
receptor implicated in fibrosis and other diseases) [104]. from archaeological records [110]. Traditional sampling
This emerging work demonstrates how deep-learning methods can be time and resource-intensive [111]. Existing
models can lead to faster and more cost‑effective applications have used pattern recognition and AI to
experiments than traditional approaches [100]. identify patterns in pottery and engraved wooden artifacts
[111, 112] as well as to sort and filter images and identify
4.2.11 Representing spatial phenomena objects in images (e.g. rock art, tools, shell or animal bone)
[110]. With advances in large-scale lidar, satellite and aerial
Geospatial AI is an emerging field that combines spatial imagery, archaeologists have access to richer geospatial
science with AI methods to derive rich insights from data for archaeological mapping of sites [110]. As a result,
big spatial data [105]. These approaches can be used to using machine learning to analyse geospatial data can help
understand the environmental factors that people may archaeologists profile the landscape characteristics without
be exposed to at a given geographical location and time physically accessing archaeological sites [110, 113, 114].
and how this exposure may impact their health [105].
A group of researchers from the University of Southern
4.2.14 Faster chip design
California used this approach to develop a model
that can predict air quality (i.e. particulate matter air When designing computer chips (such as CPUs and GPUs),
pollution <2.5 μm in diameter, or PM2.5) [106]. The model various components must be placed in a floorplan that
accurately predicted PM2.5 concentration levels without satisfies many operational requirements, including metrics
relying on prior domain knowledge and quantified the such as power consumption, performance, and chip
impact of various geographic features (e.g. parking area. This typically requires considerable manual effort
lots, commercial buildings) on air quality [106]. The fine over many months to generate manufacturable layouts.
spatiotemporal resolution of this model provided To address this problem, researchers at Google designed
insights into the impacts of air pollution (e.g. health or an AI system for automatic floorplan generation based on
environmental outcomes) on specific populations [106]. reinforcement learning (a type of machine learning which
learns from its past mistakes) that required a training dataset
4.2.12 Automating literature reviews and of 10,000 chip layouts of varying quality. The trained system
was then able to automatically generate floorplans in only
bibliometric analyses
6 hours, with the resulting floorplans having comparable or
Scientists are under increasing strain to keep up with better metrics than human‑designed floorplans [115]. Similar
the ever-growing number of scientific publications, as reinforcement‑learning approaches may be applicable
are editors and peer-reviewers [107]. Moreover, scholarly in science domains where experimental designs require
works are usually written for expert audiences in specific time‑consuming trials to explore the space of many possible
academic fields, limiting their use beyond academia, or outcomes. The potential time savings may allow researchers
even other academic fields. Emerging AI tools can assist to focus on higher level tasks and hence be more productive;
with reviewing, appraising and summarising scientific substantial time savings may also lead to discoveries
publications. Examples include the Artificial Intelligence that hitherto were too time-consuming to pursue.

21
The Healthy Country Dashboard, we developed
in collaboration with the Kakadu Rangers
and Microsoft, analyses drone footage to
autonomously identify magpie geese and
provide an accurate estimation of the area’s
bird population. Image: Microsoft

22 Artificial intelligence for science

5 Science domain
adoption trends
In this section we explore temporal patterns of AI application and development
during 63 years from 1960 to 2022 in science application domains. This shows
how the field of AI has moved beyond computer science into other scientific
and academic research disciplines. Overall, AI continues to be increasingly
adopted across all areas of science, as evidenced by the increasing share of
AI publications relative to total publications. This trend is likely to continue
for some time until AI usage is normalised within science domains and as
researchers grapple with the ever-increasing volume of scientific data.

5.1 Data sources and methods Our work contributes to this body of knowledge by using
a novel, large, comprehensive and up-to-date dataset of
(bibliometric analysis) scholarly publishing. We use a broader definition of AI with
We applied bibliometric analysis to explore trends in a larger and more diverse set of search phrases developed
AI adoption [116]. Bibliometric analysis involves the by the OECD [122] via expert consultation. Our bibliometric
examination of terms and phrases in research literature analysis is focused on AI application within other research
to understand important trends or patterns. Bibliometric fields; not the mirror (opposite) issue covered in earlier
analysis is becoming increasingly popular as the body work [120] about how other fields have been used within
of published research continues to expand [117, 118]. AI research. Our analysis uses a formal, comprehensive
The volume and rate of publishing in some fields makes and granular classification of research covering all major
comprehensive literature reviews by human researchers fields of physical, natural and social sciences and arts and
difficult or infeasible. Semi-automated literature searches humanities. We take a historical perspective examining
which augment human researchers are increasingly needed AI publication trends from 1960 to 2022. We also examine
to achieve up-to-date coverage of all relevant publications. patent citations relating to AI technologies and how
different types of AI technology have evolved over time.
Previous studies have used bibliometric analysis to explore
AI research patterns. One such study used Microsoft Our bibliometric analysis of AI research trends across
Academic Graph [119] to examine the extent to which social various scientific application domains is based on
science research fields are cited within AI publications [120]. The Lens database [123]. The Lens is a product of a
The authors found social science fields such as geography, collaboration between the Queensland University
art and philosophy were under-represented in AI research. of Technology and a not-for-profit Brisbane-based
They conclude ‘the gap between social science and AI firm Cambia. It received funding from the Bill and
research means that researchers and policymakers may Melinda Gates Foundation, the Rockefeller Foundation
be ignorant of the social, ethical and societal implications and other organisations. As of 20 September 2022,
of new AI systems’ [120]. In another bibliometric study, The Lens database contained 249 million scholarly
web-of-science data was used to examine AI research publications and over 143 million patent records.
efforts across countries, sponsors, institutions and
disciplines [121]. This study found that AI technology
development has arisen from high levels of interdisciplinary
research. TheStanford University AI index report also uses
bibliometric analysis to measure AI publishing intensity [14].

23
The Lens has previously been used to analyse science structures and definitions/rules of the database are likely
trends relating to genetics [33] and COVID-19 [124]. to be changing. This means that future extractions of
The Lens draws upon data from the following databases: the data could yield different results but are unlikely to
change the main implications/results from our study.
• CrossRef
• The Open Researcher and Contributor IDentifier (ORCID) To identify AI-related publications, we used a set of 214
• PubMed AI search phrases developed by the OECD via expert
consultation (Appendix B). AI-related publications had
• Impactstory
to contain one or more of the 214 search phrases in
• COnnecting REpositories (CORE) the publication title, abstract or keywords. We limited
• Microsoft Academic (which ceased our analysis to scholarly publications that were journal
operations on 31 December 2021) papers, books, book chapters, conference proceedings
• European Patent Office (EPO) and conference proceedings articles; all of which are
• United States Patent and Trademark Office (USPTO) peer-reviewed. This search strategy returned 3.35
million AI-related scholarly works published between
• Intellectual Property (IP) Australia
1 January 1960 and 20 September 2022 (Figure 10).
• World Intellectual Property Organization (WIPO).
Each publication was classified by field of science using the
We extracted bibliometric data from The Lens using their All Science Journal Classification (ASJC) system. The ASJC
Application Programming Interface (API) with Python is a three-level hierarchical taxonomy that represents a
scripts during 18–29 April 2022 (to capture the 1960–2021 comprehensive classification of global research covering
time period) and on 20 September 2022 (for the last, partial, all fields of study and is maintained by Elsevier (Table
year of data of 2022). The last year, therefore, contains 1). The top level has four categories (health sciences,
data for 72% of the year. Whilst the absolute numbers life sciences, physical sciences, and social sciences and
of AI publications are likely to rise and change over the humanities), the second level has 26 categories, and the
remaining days of the year, the relative shares (percentages) third level has 333 categories. The ASJC codes are assigned
are likely to remain stable. The Lens database contains by Elsevier’s team of in-house experts at the time of
records on all scholarly publications in all fields of research publication. The assignment of the code is based on the
by the whole world over all history. As such it is a large, aims, title and content of the publication [47]. A single
complex and continually evolving database. The content, publication can be assigned multiple ASJC codes.

All scholarly publications during

January 1960 to September 2022
(249,019,130)

Limited to peer-reviewed
journal papers, books, book
chapters and conference
proceedings/papers
(154,895,108)
Limited to documents with
artificial intelligence phrases
in title, abstract or keyword
(3,354,619)

Figure 10. Identifying artificial intelligence scholarly works from 1960 to 2022.
Data source: The Lens [32].

24 Artificial intelligence for science

Table 1. All Science Journal Classification (ASJC) categories [47].

First-level research field Second-level research field Number of third-level research fields
Health sciences Dentistry 7

Health professions 17

Medicine 49

Nursing 24

Veterinary 5

Life sciences Agricultural and biological sciences 12

Biochemistry genetics and molecular biology 16

Immunology and microbiology 7

Neuroscience 10

Pharmacology toxicology and pharmaceutics 6

Physical sciences Chemical engineering 9

Chemistry 8

Computer science 13

Earth and planetary sciences 14

Energy 6

Engineering 17

Environmental science 13

Materials science 9

Mathematics 15

Physics and astronomy 11

Social sciences and Arts and humanities 14

humanities
Business management and accounting 11

Decision sciences 5

Economics econometrics and finance 4

Psychology 8

Social sciences 23

25
5.2 Publishing intensity and 5.3 Adoption trends in
volumes – all research fields application domains
In this section we examine temporal patterns in AI In this section we explore AI publishing intensity within
publishing across all ASJC fields of research. We found the the four first-level (Figure 13) and 26 second-level (Table 2)
volume of AI publishing continually increased over history, research fields. We report AI publishing intensity as the
both in terms of the total number of AI publications and percentage of AI-related publications out of the total
the relative share of total publications (Figure 11). The only number of publications. We also report the AI publication
exceptions were 1964 and 1971, where the number of counts in absolute terms. Physical sciences account for
publications contracted, before returning to growth in the the bulk of AI publishing in relative and absolute terms.
following year. Interestingly, we did not find convincing In 2021, there were 461,000 AI-related publications in the
evidence of a decline in AI publications associated with physical sciences, accounting for 9.4% of total publication
1974–1980 and 1987–1993, the periods corresponding output. AI-related publishing was roughly evenly
to the first and second AI winters, respectively (Figure distributed across other first-level fields of science, with
12). However, our analysis is not exhaustive and there the social sciences and humanities, life science and health
may be time lags, spatial patterns and impacts on sciences making up 3.9%, 3.4% and 2.6%, respectively.
specialised AI fields worthy of further consideration.
We examined trends in AI publishing intensity over time to
Growth in AI publishing has been greatest in the past find evidence of AI winters across specific fields of research.
5–6 years, with the relative share of AI publishing rising There is evidence of an AI winter from 1974 to 1980 in the
from 2.9% of all publications in 2016 to 5.7% of all field of computer science (the first winter), with publishing
publications in 2022. The total number of AI publications intensity dropping and then plateauing during this period.
rose from 159,426 to 344,265 over this time period Across all other fields of research, there is no clear evidence
which equates to a 2.2-times increase. In most fields of of a similar pattern during either the first or second AI
research, the amount of AI adoption in the past several winter. This suggests the first AI winter may have been
years roughly equals what happened over all preceding isolated to the computer science domain. Beyond this we
history. Across all fields of research, the volume of did not find evidence of domain-specific slow-downs in AI
peer-reviewed publications on AI in the past 7 years publishing associated with either of the two AI winters.
(1.6 million documents) exceeds all prior AI publishing
over the proceeding 55 years (1.5 million documents). Looking across the second-level research fields, we
found that computer science dominates AI publishing,
Assuming these trends continue, it is likely that a much with one-quarter of all publications in the field on AI.
greater share of publishing will be on the topic of AI by Mathematics (14%), engineering (11%) and decision
2030. Our analysis of AI-related publications suggests that sciences (11%) also have a high AI publishing intensity.
we are currently on the steepest part of the adoption curve A similar pattern was also observed when looking at
and there are no signs of a slowdown. As such, the full the number of AI publications across fields of research.
potential impact of AI on science and research domains The lowest level of AI penetration was observed within
lies ahead. AI will become increasingly integrated into dentistry, nursing, veterinary science, pharmacology
routine research practices. As AI becomes normalised toxicology and pharmaceutics, where publishing intensity
researchers may apply AI tools and concepts without ranged from 1% to 2%. However, this appears to be
using AI phrases in the title, abstract or keywords. changing, as AI publishing has increased in these fields
(and most other fields of research) in recent years.

26 Artificial intelligence for science

SHARE OF PUBLISHING ON ARTIFICIAL INTELLIGENCE (%)
6%
2022, 5.69%

5%

4%
2017, 3.06%

3%

2%

Sept 2022
1%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 11. Peer-reviewed research publications on artificial intelligence.

Data source: The Lens [32]. Date range is from 1 January 1960 to 20 September 2022.

YEAR ON YEAR CHANGE IN AI PUBLICATIONS (%)

100%

80%

60%

40%

20%

0%
1961
1963
1965
1967
1969
1971
1973
1975
1977
1979
1981
1983
1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013
2015
2017
2019
2021

-20%

-40%

-60%

Figure 12. Annual change in peer-reviewed AI-related publications (%) over time.
Data source: The Lens [32].

27
Table 2. Artificial intelligence publishing intensity by research field (percentage).

Fields of Research (Second-Level ASJC) 1970 1980 1990 2000 2010 2015 2020 2021 2022*
Agricultural and biological sciences 0.1 0.1 0.3 0.7 1.2 1.5 2.5 2.8 3.5

Arts and humanities 0.0 0.1 0.3 0.4 0.6 0.7 2.3 3.2 2.7

Biochemistry genetics and molecular biology 0.1 0.1 0.2 0.4 1.3 1.9 3.2 3.8 4.8

Business management and accounting 0.5 0.5 0.9 1.3 2.2 2.6 4.8 5.0 6.3

Chemical engineering 0.1 0.0 0.2 0.7 1.0 1.1 4.1 4.8 5.4

Chemistry 0.0 0.1 0.2 0.4 0.6 1.0 2.7 3.2 3.6

Computer science 3.7 1.9 6.9 12.4 16.0 17.1 22.7 25.7 29.5

Decision sciences 2.3 1.4 2.1 4.5 7.1 8.5 9.8 11.3 14.9

Dentistry 0.0 0.0 0.1 0.3 0.3 0.3 0.9 1.7 1.6

Earth and planetary sciences 0.1 0.2 0.5 0.9 1.7 2.5 4.4 5.5 7.3

Economics econometrics and finance 0.0 0.1 0.3 0.8 0.9 1.1 2.7 3.5 3.9

Energy 0.1 0.2 0.3 0.8 1.5 2.1 4.5 5.2 6.0

Engineering 0.3 0.4 1.6 3.0 4.4 5.2 10.1 11.3 12.4

Environmental science 0.1 0.2 0.3 0.7 1.3 1.7 2.9 3.3 4.0

Health professions 0.1 0.2 0.4 1.1 1.4 2.2 3.2 4.1 7.3

Immunology and microbiology 0.1 0.1 0.1 0.3 0.7 1.4 1.9 2.3 3.1

Materials science 0.1 0.1 0.3 0.5 0.7 0.9 4.2 4.1 4.2

Mathematics 0.6 0.8 1.9 4.9 7.9 9.0 12.7 14.1 15.3

Medicine 0.0 0.1 0.2 0.3 0.8 1.1 2.2 2.7 3.5

Neuroscience 0.1 0.1 0.5 1.1 2.3 3.5 5.1 6.1 8.3

Nursing 0.0 0.0 0.1 0.2 0.3 0.5 1.1 1.2 1.9

Pharmacology toxicology and pharmaceutics 0.0 0.0 0.1 0.3 0.6 0.9 1.7 2.0 2.2

Physics and astronomy 0.1 0.2 0.6 0.8 1.2 1.7 5.6 7.0 7.2

Psychology 0.2 0.4 0.7 1.2 1.7 2.2 2.7 2.9 3.6

Social sciences 0.1 0.1 0.3 0.4 0.8 1.2 2.8 3.6 4.1

Veterinary 0.0 0.0 0.1 0.1 0.2 0.4 1.0 1.1 1.4

Level of AI publishing intensity Low Medium High

* Final year of data captures publishing activity within 1 January to 20 September 2022 only.

28 Artificial intelligence for science

SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

12%
Health sciences

Life sciences
10%
Physical sciences

Social sciences and humanities

8%

6%

4%

Sept 2022
2%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 13. Artificial intelligence publishing intensity by main research domains.

Data source: The Lens [32]. Data sourced from January 1960 to September 2022.

The AI publishing intensity within a field of science The development and adoption of AI technology started
is influenced by the overall publishing volume; both within the areas of computer science, mathematics,
metrics are important. For example, while medicine has engineering and decision sciences, with the AI publishing
comparatively lower AI publishing intensity relative to intensity picking up in these areas from the early 1980s
other fields (2.6% of publications were AI-related in 2021), or before. Most other fields do not show substantial
it accounts for a substantial share of AI publication volume. uptake until the 2000s or 2010s. What is consistent
In 2021 there were 55,374 AI-related publications within the across all fields of research, however, is a sudden and
field of medicine, which accounts for 8.4% of all AI-related substantial surge in AI publishing intensity from 2017 to
publications. This makes medicine the third-largest field 2021 (Figure 14 to Figure 18). Except for materials science,
by publication volume after computer science (25.5% of all publishing intensity has continued to rise, or accelerate,
AI-related publications) and engineering (18.1%). As such, in all fields of research during the COVID-19 pandemic.
medicine represents an important area for AI science Overall, AI is having a greater impact in the current era,
and technology development. Within the overall field of compared to all history, for the physical sciences, life
medicine, the usage and development of AI technologies sciences, health sciences, social sciences and humanities.
is most pronounced within the third-level fields of health
informatics, as well as radiology, nuclear medicine and
imaging, where access to digital data is readily available.

29
SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

35%
Computer science
Mathematics
30%
Engineering
Earth and planetary sciences
25%
Physics and astronomy

20%

15%

10%

Sept 2022
5%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 14. Artificial intelligence publishing intensity in the physical sciences (1–5 fields).
Data source: The Lens [32]. Data is sourced from January 1960 to September 2022.

SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

7%
Energy

Chemical engineering
6%
Materials science

Environmental science
5%
Chemistry

4%

3%

2%
Sept 2022

1%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 15. Artificial intelligence publishing intensity in the physical sciences (6–10 fields).
Data source: The Lens [32]. Data is sourced from January 1960 to September 2022.

30 Artificial intelligence for science

SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

8%
Dentistry
Health professions
7%
Medicine
Nursing
6%
Veterinary

5%

4%

3%

2%

Sept 2022
1%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 16. Artificial intelligence publishing intensity in the health sciences.

Data source: The Lens [32]. Data is sourced from January 1960 to September 2022.

SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

9%
Agricultural and biological sciences
Biochemistry genetics and molecular biology
8%
Immunology and microbiology
Neuroscience
7%
Pharmacology toxicology and pharmaceutics

6%

5%

4%

3%

2%
Sept 2022

1%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 17. Artificial intelligence publishing intensity in the life sciences.

Data source: The Lens [32]. Data is sourced from January 1960 to September 2022.

31
SHARE OF TOTAL PUBLISHING ON ARTIFICIAL INTELLIGENCE

16%
Arts and humanities
Business management and accounting
14%
Decision sciences
Economics econometrics and finance
12%
Psychology
Social sciences
10%

8%

6%

4%

Sept 2022
2%

0%
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 18. Artificial intelligence publishing intensity in the social sciences and humanities.
Data source: The Lens [32]. Data is sourced from January 1960 to September 2022.

5.4 Artificial intelligence which AI is concentrated in a few fields of research versus

evenly distributed across all fields, an approach taken
technology diffusion trends in earlier AI technology diffusion analyses [120]. The GC
In the early stages of its development AI research mostly ranges from 0 to 1, where 1 implies all AI publications
occurred within the computer sciences field and, to a lesser are in one field of research and 0 implies a perfectly
extent, engineering, mathematics and decision sciences. equal distribution of AI publishing across all fields.
However, the usefulness of AI was soon discovered by
We found that from 1960 to 1980 the GC fell from 0.9 to
scientists in many other fields of research. The pattern
0.7 as AI diffused beyond the foundation disciplines of
via which AI technologies developed by computer
computer science, mathematics and engineering into
scientists were adopted in other fields can be considered
a much broader range of application domains (Figure
technology diffusion, a sub-field of research in the
19). This level of AI diffusion has been sustained over
discipline of technology economics [125]. In this section
the following four decades. The main reason the GC
we explore the diffusion of AI technology. We do this
hasn’t fallen further, indicating a more even distribution
by analysing time-series data on the concentration of AI
of AI activity, is that computer science has increased
activity and temporal patterns of AI adoption. We analyse
AI publishing faster than other research fields.
the most granular level of the ASJC classification, the
third‑level, which contains 333 unique fields of research. The number of fields of research using AI has increased over
the same time period, with close to all fields publishing
Our analysis of the concentration of AI publishing over
on AI by 2021; up from 70% in 1980 (Figure 20). It took
time uses the Gini Coefficient (GC). The GC is a measure of
roughly 25 years from 1960 to 1985 for AI technologies
concentration for any variable across multiple categories,
to be represented in over 80% of all research fields.
such as how evenly wealth is distributed among individuals
We can see from these data that AI has been embedded
within society. We used the GC to measure the extent to
and applied in most fields of research for decades.

32 Artificial intelligence for science

LEVEL OF CONCENTRATION OF AI PUBLISHING GINI COEFFICIENT

1.00

0.95

0.90

0.85

0.80

0.75

0.70

0.65

0.60

Sept 2022
0.55

0.50
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 19. Concentration of artificial intelligence publishing across research fields.

Note: The Gini Coefficient is a statistical measure of concentration. It is used here to measure the level of concentration
of AI publishing across fields of research. A higher value represents increased concentration.

NUMBER OF RESEARCH FIELDS WITH AI PUBLISHING (TOTAL = 333 FIELDS)

350
Total number of research fields (333)

300

250

200

150

100
Sept 2022

50

0
1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Figure 20. Diffusion of artificial intelligence technology into research fields.

Note: There are a total of 333 third-level All Science Journal Classification research fields. The graph shows
the number of research fields with artificial intelligence publishing from 1960 to 2021.

33
5.5 Trends in artificial We also examined the extent to which various AI
technologies have been translated into commercially
intelligence technologies valuable and socially useful products. We did this by
In this section we explore publishing trends across various examining the number of patent citations attributed
subfields of AI. We analysed publication counts for the to research publications containing an AI-related
214 OECD AI phrases (Appendix B) from 1960 to 2021. search term in the title, abstract or keyword. Patent
The majority of AI technologies saw publishing volumes citations were taken as a proxy measure of commercial
increase over time, with some developing faster than (and societal) impact, where a greater number of
others. The top six AI technologies with strongest growth patent citations is assumed to indicate greater impact.
over the past 20 years were convolution neural networks, The analysis revealed that scholarly publications which
followed by deep learning, random forest, generative included reference to ‘neural network’, ‘convolutional
adversarial networks (defined as ‘adversarial network’), neural network’, ‘deep learning’, ‘machine learning’
sentiment analysis and transfer learning (Figure 21). and ‘computer vision’ generated the greatest number
Deep learning accounts for the greatest share of AI of patents (Figure 22). These AI subfields are associated
publications in 2021 (7.6% of total AI publications) and has with the most significant commercial value and impact.
shown the strongest 20-year increase, growing 25.4 times
from 2002 to 2021. Looking at the past 5 years alone, the
strongest growth has been observed across ‘generative
adversarial networks’ and ‘transfer learning’ (growing
by 2.3 and 5.3 times during 2016–2021, respectively).

THOUSANDS OF PUBLICATIONS

60
Adversarial network

Sentiment analysis
50
Convolutional neural network

Random forest
40
Transfer learning

Deep learning
30

20

10

0
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021

Figure 21. The six fastest-growth artificial intelligence technologies over past 20 years.
Data source: The Lens [32], OECD [122]. Note: The graph shows the number of peer-reviewed publications with the search term
appearing in the title, abstract or keywords for journal papers, conference papers, books and book chapters.

34 Artificial intelligence for science

Adversarial network 830

Artificial intelligence 1,252

Computer vision 2,350

Convolutional neural network 5,276

Deep convolutional neural network 1,421

Deep learning 7,679

Deep neural network 1,982

Feature extraction 1,220

Image classification 1,127

Image processing 1,125

Image segmentation 865

Machine learning 5,824

Neural network 11,265

Object detection 1,887

Reinforcement learning 999

Robot 1,797

Speech recognition 1,173

Support vector machine 937

Transfer learning 868

Vector machine 946

Figure 22. Patent citation counts for the top 20 artificial intelligence phrases for 2017–2021.
Data source: The Lens [32].

35
Year 6 students from a Sydney
public school test out the
Smart Bin’s recycling skills.
Image: Keith McInnes Photography

36 Artificial intelligence for science

6 Future development pathways
6.1 Software, hardware and state-of-the-art classical supercomputer 10,000 years to
solve. They refer to this as an ‘experimental realisation of
open access resources quantum supremacy’ [127]. The researchers conclude that
Hardware and software toolkits available to scientists ‘as a result of these developments, quantum computing
and researchers wanting to apply AI in their work are is transitioning from a research topic to a technology
continually improving. This will boost the productivity of that unlocks new computational capabilities’ and that
advanced AI developers and researchers with limited AI ‘we are only one creative algorithm away from valuable
knowledge seeking to perform a particular AI function near-term applications’ [127]. Quantum computers may
or procedure. The technology upgrades are having eventually have the capacity to solve AI problems beyond
the impact of democratising and industrialising AI. the reach of conventional computers. This can potentially
They will play an important role in the diffusion of AI lead to a paradigm shift and step change in AI capability.
technology across all fields of science and research.
In addition to improved hardware, the AI field is seeing
AI accelerators are computing processors specifically the rapid growth of software frameworks to support AI
designed to handle matrix algebra operations used in operations. Examples of popular frameworks include
machine learning. AI accelerators improve the speed PyTorch, Tensorflow, Keras and Caffe. Using these
and reduce the latency of AI computations. They enable frameworks in environments such as Python and R,
more time-efficient and cost-efficient development of researchers can design and/or adapt machine-learning
AI systems. Over the past few years, the number of AI algorithms relatively quickly, often without the need to
accelerators available on the market has increased in delve into the low-level details of the algorithms. This is
quantity and diversity. They are used in applications such how many researchers are likely to develop and apply
as autonomous vehicles, speech recognition, natural AI within their fields of expertise. These AI frameworks
language processing and video object detection [126]. have played, and will continue to play, an important role
A recent review [126] of AI hardware accelerators by the in facilitating AI technology diffusion across all fields of
Massachusetts Institute of Technology describes over 70 AI physical, natural and social sciences. We are also seeing
accelerators under several categories: (a) research chips, the emergence of code-free AI software tools delivered
(b) very low power chips, (c) embedded chips and systems, through graphic user interfaces (GUIs). A research team
(d) autonomous systems, (e) data centre systems, (f) data from Moorfields Eye Hospital in the United Kingdom
centre chips and cards. The last category is further broken recently evaluated code-free AI tools for training
up into CPUs (with new low-level instructions targeting machine-learning algorithms from corporations such as
AI/ML workloads), FPGA-based accelerators, GPU-based Amazon, Apple, Clarifai, MedicMind and Microsoft [128].
accelerators, and dataflow chips, such as Google’s Tensor The code-free deep learning (CFDL) software tools
Processing Units (TPUs). In addition to those already on were used for the classification of medical imagery.
the market, the study reviews 12 AI accelerators that have They conclude ‘that CFDL platforms have the potential
been announced for future release in the short term, to improve access to deep learning for both clinicians
including accelerators from Qualcomm, a large company and biomedical researchers, and represent another
in the mobile/cellular phone space. The future is likely step towards the democratization and industrialization
to see continued improvement in computational power of AI’ [128]. Furthermore, mass-market software tools
and efficiency of AI systems with an increased diversity of such as Microsoft Excel and Microsoft Power BI are
specialised processors suited for various applications. increasingly making code-free machine-learning functions
available to users to perform common tasks [129, 130].
A related development in computing hardware is the
rise of quantum computing. Quantum computers use The open-access frameworks for AI computations are
theories of quantum physics to store and analyse data. supported by a large and growing number of platforms
Unlike conventional computers, which use a binary (on/ which enable knowledge sharing. Examples include
off) system to represent data, quantum computers use GitHub, Bitbucket, SourceForge, Gogs, Gitbucket, AWS
qubits which can be in many states at any given point CodeCommit, Beanstalk, Phabricator, Gitea, Allura,
in time. Writing in Nature in 2019, Google scientists Rhodecode, CodeGiant, Cloud Source Repositories
reported that their quantum processor called ‘sycamore’ (by Google), Azure DevOps Services, Google Developers
can solve a problem in 200 seconds that would take a and Trac [131]. There are many other such platforms.

37
These are powerful information resources which speed 6.2 The quest for better data
up and assist scientists developing software code for
AI. The volume of material and continuously improving We are living in the era of ‘big data’, where the volume,
search tools allow a software developer to find a code variety and velocity of data inflows continue to expand.
snippet, library or dataset to quickly solve a problem Big data have supported the training of machine‑learning
they’re working on. These platforms also facilitate algorithms. For example, vast image datasets (labelled by
Q&A style discussions where software developers can users through search terms) supported the development
turn to their community for help. Furthermore, there of image recognition systems able to accurately
are platforms such as Kaggle and ImageNet which host identify dogs, cats, birds or practically any object
competitions for AI experts to access datasets and solve within an image. Speech recognition, face recognition
problems. Competitions on these platforms have fast and emotion/expression recognition systems have
tracked AI problem solving in many areas [132, 133]. benefited similarly from vast volumes of labelled data.
Collectively these open-access resources will provide a
However, big data can be problematic. Big data contain
big boost to AI application in diverse fields of science.
considerable noise in addition to the signal. Big data
Lastly, the rise of accessible cloud-based computing can contain spurious entries which are camouflaged and
services is also facilitating the adoption of AI across all hard to identify amid the other entries. This can degrade
fields of research and world regions. A recent report the accuracy and reliability of machine-learning models.
by information technology (IT) consulting firm Gartner For example, a recent analysis of 62 published scientific
finds that the global cloud computing market grew from studies using machine learning on chest radiographs and
US$270 billion in 2020 to a (forecast) US$397 billion in CT scans to detect and prognosticate COVID-19 found that
2022 with 23% growth during 2021 [134]. The analysts ‘none of the models identified are of potential clinical use
observed that the pandemic fuelled the growth of cloud due to methodological flaws and/or underlying biases’
computing with many business operations moving online. [7]. Most of the problems related to duplication and
A similar pattern is likely to have occurred within the global quality issues in the datasets used for machine learning:
science and research community as remote work was
• Incorrectly sourced datasets – In these cases
needed due to movement and/or quarantine restrictions.
data were incorrectly sourced from demographic
Market research by ReportLinker forecasts the continued
age groups that led to biased and inaccurate
growth of the cloud rising at an annual compound annual
results when applied at the population level.
growth rate of 16.3% during 2021 to 2026 and reaching
• Frankenstein datasets – In these cases public datasets
US$948 billion per year [135]. The science sector is an
were assembled from numerous other datasets and
enthusiastic adopter of cloud computing [136]. The field
then redistributed under a new name. This meant
of genetics, for example, depends on cloud computing for
algorithms were being trained on multiple identical
storage, sharing and analysis of vast quantities of data for
or overlapping datasets with significant duplication.
cross‑organisational and international science teams [137].
• Biased datasets – In these cases images shared
publicly and/or contained within published
6.1.1 Implications for science and documents often have a form of selection bias.
research organisations For example, people with certain conditions and/
or disease severity may be more/less likely to share
Science organisations seeking to uplift AI capability will
their images. This leads to bias in the data.
need to make decisions about hardware, software and
computational infrastructure upgrades, including the access The masses of data, which on the surface looked like a
to cloud computing services. These tools have improved powerful resource for training machine-learning algorithms,
substantially over recent years. They are likely to follow had serious limitations which in every case made the
a pathway of ongoing improvement in the future. There model unusable in clinical settings. Similar problems have
are many unknowns about quantum computing; it could been observed in earlier reviews of AI-based models for
potentially lead to a step change and paradigm shift in COVID-19 diagnosis and prognosis. For example, another
AI resulting in substantially elevated capability. Quantum review of 169 studies containing 232 machine‑learning
computing services are already available to AI developers models found ‘all models were rated at high or unclear
and science organisations will need to factor this into risk of bias, mostly because of non-representative
their longer-term AI capability development strategies. selection of control patients, exclusion of patients who

38 Artificial intelligence for science

had not experienced the event of interest by the end of on the right (top-priority) datasets. Some of the
the study, high risk of model overfitting, and unclear datasets needed will be novel and will require new
reporting’ [138]. Most of these problems stemmed from the investment/capture. Other datasets will be historic and
incorrect use of datasets that were not fit-for-purpose. may require formatting, validating and fixing. The data
imperative means that research organisations, like any
Given such findings the AI science community in organisation, need to move towards becoming increasingly
healthcare, and other domains, is likely to invest greater data‑driven. This implies changes to business processes,
effort in developing higher quality and fit-for-purpose infrastructure, skills and organisational culture.
datasets. Professor Luciano Floridi from the University
of Oxford and Alan Turing Institute recently published Strategies about how to become a data-driven organisation
a paper [139] on the near-term future for AI. One of the are well covered in the management sciences literature
trends he identifies is the move from ‘big data’ to ‘small [143, 144]. Data-driven organisations need strong
data’ – and he defines small data as being higher quality, capacity to acquire, analyse, interpret protect, store,
well curated and provenance assured. He gives the share and communicate data. Furthermore, data-driven
example of an AI-based system developed by Google’s organisations demonstrably use data in decision making
DeepMind in partnership with Moorfields Eye Hospital to achieve organisational objectives. They also know the
in London. Historically, medical imaging diagnostics value of their current and future-planned data assets.
based on AI typically relied on ‘databases of millions of Research organisations need to acquire these traits
annotated images’. However, the system was successfully to achieve aspirations for AI capability upgrades.
trained using only 14,884 eye scans for early detection
of sight‑threatening eye diseases and was found to
reach or exceed diagnostic accuracy by experts and is 6.3 Education, training
considered clinically applicable [140]. The dataset of and capability uplift
14,884 scans represents a considerably smaller‑than‑usual
The surge in AI development and application is being
dataset. The dataset wasn’t just smaller; it was also well
accompanied by a surge in AI training and education.
curated, labelled by experts, reviewed/examined by
An analysis of 18 universities across 9 countries found
experts, provenance assured, and fit-for-purpose [140].
that the number of undergraduate courses teaching
students skills necessary to build and deploy AI models
6.2.1 Implications for science and doubled from 2016 to 2020, and increased by 42%
research organisations for postgraduate courses [14]. Similarly, enrolments
in introductory courses for AI or machine learning
The ability of a research organisation to achieve have grown by close to 60% over the same time [14].
competitive differentiation and problem-solve with AI Data from the OECD AI Policy Observatory shows the
will, in large part, be determined by the quality of its number of AI courses (delivered in English) worldwide
datasets. Vast volumes of publicly available data have increased 80.1% during 2018–2021, and AI now comprises
been, and will remain, important for the development 27.3% of all computer science and IT courses [145].
of AI. For example, a CSIRO team used machine learning
to identify COVID-19 virus mutations with likely impact In Australia the number of AI courses offered by universities
on disease severity [141]. This was done on only 0.3% of has grown 1.2 times over the past 4 years, with 235 courses
the available viral data points due to a lack of patient on offer in 2021 [145]. The University of Queensland and
information for the rest [142]. This illustrates the rate- the University of Sydney are ranked among the world’s top
changing potential for high-quality datasets. Specifically, 100 academic institutions in AI according to the Nature
future AI capability uplift will require investments in high- Artificial Intelligence Index. These institutes are placed in
quality data which is fit-for-purpose, provenance assured, 55th and 76th positions, respectively [146]. The universities
validated, up-to-date, and ethically obtained. As such, it ranked in the top 10 positions are all in the United States,
is likely to be critical for an organisation to know what Germany and the United Kingdom [146]. Australian
datasets it owns, their provenance, reliability and suitable institutions such as the Australian Institute for Machine
uses, as well as detailed metadata of those datasets. Learning at the University of Adelaide are also expanding,
with the number of staff increasing from 80 in 2017 to 140
This means science and research organisations seeking in 2021 [147]. Tertiary educational and vocational training
to upgrade AI capability will need to become adept at institutes in Australia are offering a growing range of
acquiring, storing, protecting and recording metadata studying opportunities for people seeking to gain AI skills.

39
In Australia the fastest growth in AI course offerings was of IT graduates in this sector grew over 5-times between
observed for master’s degrees [148]. The science and 2013 and 2020 to over 15,600 students [153]. This can be
research sectors are in direct competition with industry for partly attributed to the COVID-19 pandemic as IT degrees
AI skills. An analysis of job postings in the United States in may have been more easily shifted to online education.
2019 found 9.1% of postings were for AI-related positions
[149]. These AI-related jobs tended to be higher skilled Research has shown that workers who have a background
positions with 80% requiring a 4-year bachelor degree as a in computer science and programming are more supportive
minimum requirement [149]. In Australia, 37,587 AI-related of the development of AI [155], suggesting that AI literacy
job advertisements were posted in 2015–2019 by Adzuna – a can be an additional important factor in encouraging
leading job advertisement search engine [150]. These jobs adoption of AI across the science sector. While there is
tend to be concentrated in Australian states with the largest limited understanding around the level of AI awareness
population, with the notable exception of the Northern and understanding across the science sector in Australia,
Territory, which had a higher rate of AI jobs relative to the majority of the Australian public report low (62%) or
its population. AI-related positions make up a fraction of moderate (26%) subjective knowledge of AI [156]. Examining
the total job advertisements in Australia, accounting for and improving the AI literacy of the workforce will help
0.5% of all postings made between 2015 and 2019 [150]. science and research organisations identify gaps in the
current understanding of AI in the scientific workforce and
An analysis conducted by the OECD examined the future capability areas that require attention [155, 156].
prevalence of AI skills across occupations using LinkedIn
member profiles in 2015–2020 [151]. This analysis found Another dimension to AI upskilling is the importance
Australia ranked in the middle of the list (13th place out of of interdisciplinarity [157]. Research projects using AI
26 OECD countries) with the highest penetration of AI skills typically require high levels of interdisciplinarity involving
in the United States, followed by Germany and Israel [151]. expertise in the science application domain along with
When it comes to the AI scientific workforce, as measured specialised expertise in areas such as machine learning,
through the Global AI Talent Tracker by MacroPolo, the natural language processing, computer vision, robotics
majority of AI scientists are currently based in the United and other sub-fields of AI. A study into the interdisciplinary
States (59%), followed by China (11%) and Europe (10%) nature of AI finds that ‘the relationship between AI and
[152]. Most AI scientists completed their undergraduate interdisciplinary research must be considered as a two-way
degrees in China (29%), followed by the United States street’ [157]. The authors of this study note that more effort
(20%) and Europe (18%). This reveals a strong net-inward is going in one direction (applying AI to other research
movement of AI talent into the United States [152]. fields) than the other (applying other research fields to
AI). Other researchers using bibliometric analyses have
A growing number of students in Australia are graduating observed similar patterns [120]. Both these studies identify a
with degrees in IT or computer science and have skills need for improved two-way interdisciplinarity collaboration
suitable for AI-related occupations. The number of to achieve improved outcomes from AI for science.
university graduates with degrees in IT, as a field of
education, has been on a steep rise since 2017 after over a Lastly, it is worth noting that the future AI talent pipeline
decade of slump [153]. Between 2003 and 2013 the number for science and research organisations does not start at
of IT graduates was mainly in decline, which was a warning university. Education researchers have found an interest,
trend for Australia’s transition to the digital economy [154]. motivation and capability for science, technology,
In 2013, the trend reversed to a slow growth (2013–2017) engineering and mathematics (STEM) expertise – including
and moved into a steep rise since 2017 [153]. By 2020, mathematics and the foundational skills needed for
the number of IT graduates in Australia almost tripled AI – is typically acquired in early childhood learning
compared to 2013, exceeding 31,700 people [153]. In 2020, IT [158, 159], primary school and high school [160, 161].
became the fourth-largest field of education by the number Therefore, a longer-term view of the AI talent pipeline
of graduates after management and commerce, society requires investment in all lifelong stages of learning.
and culture and health [153]. The growth was the highest
among the postgraduate overseas students – the number

40 Artificial intelligence for science

WANDA uses AI to identify different species of fish as they’re caught in real time. We developed this new technology
to support sustainable fisheries management and seafood supply chains.

6.3.1 Implications for science and and education, research organisations will need strategies
research organisations to bolster two-way interdisciplinary collaboration in AI
projects. This involves a flow of expertise from AI specialists
Research organisations will need strategies for talent into science application domains along with the flow
acquisition and retention, given the strong demand for of science domain expertise back into the field of AI.
AI skills. Industry employers are often able to lure skilled
AI workers with high salaries. Research organisations Some of the upskilling required will exist beyond
will need to offer competitive renumeration packages the organisation’s immediate sphere of influence.
to attract and retain skilled AI workers. The feasibility of Longer term and broad-based development of the AI
talent acquisition and retention will need consideration talent pipeline will involve developing foundational
by research organisations while making decisions skills in mathematics and computational logic for
about whether to grow certain types of AI capability. children and teenagers in early learning, primary school
At some point the supply of AI skills is likely to adjust and high school contexts. That’s when an interest,
to meet demand. Many industries are on the same motivation and capability for advanced STEM skills
steep AI adoption curves as we have shown for the begins to develop. Research organisations can work
science sector; they are hungry for the same AI skills. with schools and learning organisations to promote
the foundational education needed for the future AI
The number of, and variety of, education and training workforce. In the same way that science and research
courses available to scientists seeking to upgrade AI organisations have developed school engagement
capability is continuing to expand. There is a wide range of programs to raise awareness around STEM careers
course formats: from micro-credentialling or flash courses (e.g. CSIRO’s STEM Professionals in Schools program),
to acquire specific skills for specific tasks and timeframes, similar efforts can be used to strengthen the knowledge
through to longer in-depth courses designed to develop and understanding around AI-related career opportunities.
deeper skills and knowledge. Research organisations,
and researchers, can take advantage of these educational There is also value in uplifting societal awareness and
offerings to upgrade AI capabilities. Furthermore, research understanding of AI. Improving the general knowledge
organisations may already have staff with professional of AI will help create informed users who can better
backgrounds which make them well-suited for a career manage the risks, and harness the opportunities,
transition into AI-focused roles. In addition to training associated with AI technology. It will also help society
work towards effective policies, laws and regulations.

41
6.4 Towards collaborative This AI system, developed by Facebook and NYU School
of Medicine’s Department of Radiology, interprets lower
artificial intelligence quality image data, which has been rapidly acquired, and
Initial predictions about the impact of technology on predicts the missing data to create a higher quality image.
the workforce focused on the areas where technologies This higher resolution image can then be interpreted by
like AI would substitute and replace humans [162, 163]. the doctor to determine whether an abnormality is present,
However, increasingly the focus has shifted to how increasing their productivity, reducing patient time in the
humans and AI can work together (augmentation). MRI scanner and potentially increasing diagnostic accuracy.
Human-AI collaboration is a field of research about how
Another emerging area of research related to human‑AI
humans and AI can meaningfully interact and cooperate
collaboration looks at how workers perceive future
to carry out tasks to higher standards than either can
AI developments. Initial surveys and interviews of
achieve alone [164]. Human-AI collaboration can lead to
healthcare professionals, librarians and qualitative
significant productivity gains and expand the bounds of
researchers, data scientists and the public have revealed
human capacity. IT company Accenture estimated that
a number of common themes around how workers view
organisations that invested in human-AI collaboration
human‑AI collaboration [155, 156, 167-178]. First, there is
would increase their revenues by 38% and employment by
a generally positive view towards the value that AI can
10% between 2018–2022 [165]. Human-AI collaboration is
provide. A 2019 survey of the American public found
important within the science sector as per other industry
that 79% of participants were either supportive or
sectors. The productive use of AI by scientists depends
neutral towards future AI developments, with support
in large part upon the quality of human-AI collaboration
for AI strongest in higher educated and higher income
and individual, team and organisational levels.
cohorts, or those that have a computer science or
How does human-AI collaboration happen? Partnership on programming background [155]. This share is even
AI, a global multistakeholder organisation, has developed higher in similar surveys that have been conducted in
a standardised framework to investigate and characterise Australia (85% supportive or neutral towards AI) [178].
human-AI collaboration [166]. Using seven case studies,
The positive sentiment is driven by the perceived benefits
the framework covers the following: the nature of the
of AI. A survey of clinicians in South Korea found that
collaboration (e.g. the goals of the interaction, how the
83.4% felt AI would be useful in medicine, particularly
human and AI are engaged and their level of agency); the
for diagnostic purposes [179]. The European Society of
situational context of the collaboration (e.g. whether the
Radiology also found that radiologists felt that AI can
human and AI are physically co-located, AI awareness, trust
potentially result in higher productivity, resulting in more
in AI system and potential consequences); the AI system
available time to spend with patients [168]. AI also opens
characteristics (e.g. whether the AI system is interactive,
opportunities for researchers to access and derive greater
adaptable, predictable, explainable and human-like);
value from existing large datasets that would otherwise
and the characteristics of the human collaborator [166].
be prohibitively resource-intensive to manually analyse
Understanding and examining the nature of human-AI
[170-172]. Scientists traditionally use meta-analyses to
collaboration is critical to progressing the responsible
synthesise findings from a large collection of scientific
design and governance required to ensure safe, reliable and
studies, a highly time and labour-intensive process. AI could
productive design and development of collaborative AI.
provide a means to automate the process [171], helping
FastMRI is an example of a human-AI collaborative system researchers stay across the fast-moving research landscape.
which aims to accelerate the rate at which doctors can
Studies exploring perceptions of AI in the research
acquire brain scans using magnetic resonance imaging
community have found that most scientists and researchers
(MRI) without compromising on the image quality [166].
do not think AI can, nor should, replicate the research
process [169, 172]. Instead, AI systems and human scientists
could operate in a ‘synergistic partnership’ [171-173, 180].

42 Artificial intelligence for science

This synergistic relationship acknowledges the unique 6.4.1 Implications for science and
and complementary strengths of humans and AI systems research organisations
and provides opportunities to collaborate to address the
limitations of humans and AI on their own [167, 169-172, The ability to capture value of human-AI collaboration
175, 179-181]. Under a human-AI collaboration scenario, rests upon buy-in from the scientific community and
the human scientist would delegate tasks that can be there are several factors that influence this. One of
completed more efficiently by an AI system, leaving the these is trust, particularly in non-data science domains,
human scientist to invest their time and resources into which relates to AI literacy. Research has shown that
tasks that rely on uniquely human cognitive abilities. workers who have a background in computer science and
programming are more supportive of the development
AI systems have an advantage over human workers when of AI [155]. While there is limited understanding around
it comes to rapidly processing and analysing large masses the level of AI awareness and understanding across the
of information and identifying patterns or relationships science sector in Australia, the majority of the Australian
[171, 180]. By automating these manual tasks, scientists public report low (62%) or moderate (26%) subjective
can have more time for creative and higher order knowledge of AI [156]. Examining and improving the AI
tasks [170, 171]. Conversely, humans perform better in literacy of the workforce will help science and research
ambiguous or uncertain decision-making contexts [180]. organisations build understanding and acceptance of
Researchers are sceptical whether an AI system will be AI [155, 156]. In general, a greater share of Australians
able to replicate a human’s ability to make complex places their trust in AI systems (41%) than other countries
associations or complete tasks that require subjective (e.g. 35% in the United Kingdom and 33% in the United
judgement or specialised knowledge [172]. This includes States) [156]. To maximise the complementary strengths
tasks that require prior knowledge around the data or of human scientists and AI systems, AI applications need
domain [169] or those that require human judgement, to be designed and evaluated with the human scientist
such as reviewing scientific publications [169-172] or workflow in mind. In certain cases, AI systems might be
interpreting ambiguous medical results [179, 182]. used to automate a discrete part of the workflow (e.g.
pre-processing large datasets), or work in concert with
Surveys of scientists have found that transparency and
the human scientists (e.g. validating previous analyses).
explainability are critical factors in determining trust in
an AI system [169]. Without this transparency, researchers Addressing the productivity slump is an area of concern
are concerned that AI could produce biased results or in the contemporary research sector. While AI presents
exacerbate existing societal inequalities [169]. A group of an opportunity to do more with fewer resources and
researchers from IBM and Rensselaer Polytechnic Institute free scientists up for higher value tasks, it is important
have shown that ‘transparency features’ are critical that this increase in productivity does not coincide with
in human-AI systems as they help build trust between a decline in research quality or impact [171]. For example,
the user and the machine [183]. These researchers used using AI might speed up the rate at which scientists can
visualisations to enhance transparency, and in turn, user collect and analyse data, but this may not contribute
trust, around the data that goes into the model and the to meaningful advancements in the field if such tools
process through which the AI generates a predictive model. are implemented in the absence of sufficient human
oversight and specialised domain input. Multidisciplinary
research teams involving data science and domain
knowledge specialists will likely feature more heavily to
support quality outputs from human-AI collaborations.

43
6.5 Artificial intelligence In addition to gender issues the STEM workforce, likely to
reflect the AI workforce, has a lack of cultural and ethnic
workforce diversity diversity. For example, professional body Science and
There is a lack of gender and cultural diversity in the AI Technology Australia finds that ‘one in 200 Aboriginal or
research workforce. This is evident in both technology Torres Strait Islander people of working age have a STEM
corporations and academic/research organisations. degree – while one in 20 non-Indigenous working age
For example, a recent study of gender diversity in AI people have a STEM degree’ [186]. Australian mathematician
research workforces based on analysis of arXiv publications Rowena Ball writing on the topic in Australian Quarterly
by the National Endowment for Science, Technology in 2015 says that ‘unless this percentage of Indigenous
and the Arts in the United Kingdom found [184]: enrolments in STEM is increased what does follow is
that Indigenous people are being systematically locked
• Worldwide, 13.8% of authors on AI research papers are out of high paying jobs in science related f­ ields’ [187].
female and the portion of papers written by at least It also means that Australian science and technology
one female author has not increased since 1990. is not capturing the full benefits of Indigenous science
• Less than 25% of AI researchers are female in most and knowledge. This knowledge can help us understand
academic institutions with a few exceptions. the world and problem-solve in many contexts.
• There are relatively few female authors of AI research
papers from technology corporations such as Google
6.5.1 Implications for science and
(11.3%), Microsoft (11.95%) and IBM (15.7%).
research organisations
• AI research papers with at least one female author
tended to be more applied and were more likely Many research organisations acknowledge the lack of
to use human terms such as ‘fairness, human gender and ethnic/cultural diversity in AI and STEM
mobility, mental, health, gender and personality’. workforces as a challenge they are working to address.
While there has been some progress, much remains
Gender disparity in the AI workforce is similar to the to be done. Improving workforce diversity will be an
gender disparity in the broader field of computer science important developmental pathway for AI capability uplift
and within STEM disciplines. According to the 2021 STEM within research organisations over the coming decade.
equity monitor by the Australian Government [185], A review of Australia’s strategies to achieve gender
28% of the STEM workforce are women, and males in equality in STEM was recently published by authors from
STEM professions earn (on average) $28,994 per year several universities and non-government organisations
more than females compared to a pay gap of (NGOs) [188]. This provides details on the outcomes
$25,534 across all industries. The STEM equity monitor associated with various strategies and priorities for the
also records changes to gender disparity over time: journey ahead. There are also initiatives to promote
Indigenous science in Australia. For example, the national
• During 2015 to 2019 the proportion of women
science agency of Australia, CSIRO, has an Indigenous
enrolled in STEM courses at Australian
science program which aims ‘to create Indigenous-
universities increased from 34% to 36%.
driven science solutions that support sustainable futures
• During 2016 to 2020 the proportion of
for Indigenous peoples, cultures and Country’ [189].
women working across all STEM-qualified
In one project under this program Microsoft, CSIRO
industries increased from 24% to 28%.
and Kakadu National Park rangers are combining AI,
• During 2016 to 2020 the proportion of managers science and Indigenous knowledge for environmental
and senior managers in STEM roles who are management and biodiversity protection [190].
female increased from 18% to 23%.

44 Artificial intelligence for science

6.6 The rise of ethical • Contestability: When an AI system significantly
impacts a person, community, group or environment,
expectations and regulations there should be a timely process to allow people to
A recent review of AI ethics policies was published in challenge the use or outcomes of the AI system.
January 2021 by researchers at the School of Public Policy • Accountability: People responsible for the various phases
at the Georgia Institute of Technology [191]. They identified of the AI system lifecycle should be identifiable and
112 documents prescribing AI ethics principles, frameworks, accountable for the outcomes of the AI systems, and
policies and strategies from 25 countries produced human oversight of AI systems should be enabled.’
during 2016–2019. The documents were published by
governments, companies and NGOs. The top five (of 25) These principles are identified by the Australian
ethics topics covered in these documents were: (a) social Government as voluntary and intended to be
responsibility; (b) transparency; (c) bias and fairness; (d) ‘aspirational and to complement – not substitute –
privacy; and (e) safety and reliability. The authors found existing AI regulations and practices’ [195]. A recent
that ‘public and NGO documents are more participatory review of the application of these principles
in their creation and more engaged with the law’ and was done by researchers at CSIRO [196].
that ‘private-sector documents appear to be more
As we look into the future and across the globe, it is
concerned with client and customer-related ethical
possible that the currently voluntary and aspirational AI
issues that may lend themselves to a technical fix’.
principles may become regulations and laws. A recent
Overall, the study points towards a substantial expansion April 2021 paper in the Harvard Business Review [197]
in AI ethics expectations across all sectors with the public explores this issue and opens with the statement
and NGO sectors leaning towards future legislative ‘Over the last few weeks, regulators and lawmakers
implications. It complements several earlier studies around the world have made one thing clear: New
examining the development of AI ethics policies, laws, laws will soon shape how companies use artificial
guidelines and frameworks across the globe [192-194]. intelligence’. Examples of recent developments include:
One study found convergence around five ethical principles
On 31 March 2021 the five main financial regulators
across the globe: (a) transparency; (b) justice and fairness;
in the United States (including the Federal Treasury)
(c) non-maleficence; (d) responsibility; and (e) privacy
issued an information request to financial institutions
[194]. All these topics feature in the AI ethics principles
to provide detailed information on their use
of the Australian Government, quoted as follows [195]:
of AI and machine learning. They indicated the
• ‘Human, societal and environmental wellbeing: information provided is to help ensure ‘compliance
AI systems should benefit individuals, with applicable laws and regulations’ [198].
society and the environment.
On 21 April 2021 the European Union proposed the
• Human-centred values: AI systems should respect human first legal framework on AI which includes fines of
rights, diversity, and the autonomy of individuals. up to 6% of company revenue for non-compliance
• Fairness: AI systems should be inclusive and accessible, [197, 199]. Furthermore, the European Union’s general
and should not involve or result in unfair discrimination data protection regulation (GDPR) includes articles
against individuals, communities or groups. limiting the use of automated decision systems including
• Privacy protection and security: AI systems should requirements related to explainability and contestability.
respect and uphold privacy rights and data
Clearly there is a considerable pathway ahead before
protection, and ensure the security of data.
AI-specific laws are enacted across the globe. Sectors
• Reliability and safety: AI systems should reliably such as finance and retail may be at the forefront of
operate in accordance with their intended purpose. these regulations due to their extensive and routine
• Transparency and explainability: There should handling of confidential customer data. However, over
be transparency and responsible disclosure so the coming years and decades AI policies, regulations
people can understand when they are being and laws are likely to increase. The science sector will be
significantly impacted by AI, and can find out impacted along with other sectors; research organisations
when an AI system is engaging with them. will need to ensure they are compliant. Furthermore,
AI ethics go beyond compliance. There are also rising
expectations for ethical AI from society, investors
and AI researchers and developers themselves.

45
We have worked with a range of stakeholders,
including Google and the international Kaggle
community, to explore ways to help with the
monitoring and detection of crown-of-thorns
starfish using machine learning.

However, ensuring that the development and application in the future. Societal awareness about the issues and
of AI is both compliant and meets (and exceeds) the ethical expectations for ethical AI is likely to rise. Over the last
expectations of society can be challenging. The AI research several years, governments, companies and not-for-profits
community is still working to resolve the operational have identified principles and expectations for ethical AI.
meaning of concepts such as explainability, transparency, There are high levels of agreement in these principles
repeatability and interpretability as they apply to about transparency, fairness, explainability and privacy.
machine‑learning systems. There is also considerable work
underway to develop software and systems to deliver on However, merely signalling an intention to deliver
AI ethics. For example, a recent study [200] identifies and ethical AI may not be sufficient. Delivering on complex
reviews state-of-the-art technologies to enable explainable ethical requirements will require improved scientific
AI (XAI), including: (a) features-oriented methods, (b) global knowledge and technological capability. It will require
methods, (c) concept models, (d) surrogate models, (e) skills and capability uplift within the AI workforce.
local pixel-based methods, and (f) human-centric methods. Early investment in ethical capability – including
Another area of technological innovation to achieve technology, skills and cultures – will help research
improved AI ethics is privacy-preserving analytics. Recent organisations stay ahead of the regulations.
review papers have been published on this rapidly emerging Lastly, there’s a complex balance between efforts to ensure
field [201, 202]. There is also a growing body of work and the ethics of AI and the development of novel technologies
technology development on improved ways to identify which improve (or save) people’s lives. Effective approaches
and manage bias in machine-learning projects; a recent to AI ethics will ensure principles are upheld without
review paper describes 25 bias mitigation methods [203]. limiting the pace or quality of innovation and discovery.
Furthermore, many of today’s innovative technologies and
approaches enabling improved ethical performance – as
6.6.1 Implications for science and
discussed above – have grown organically within the AI
research organisations
community. This has mostly happened in the absence of
The implication arising from this AI development pathway is laws and regulations. There’s much evidence of a strong
that the AI ethics performance bar is likely to be higher and drive, coming from within the AI research and development
more tightly regulated into the future. What are currently community itself, to achieve improved ethical performance.
voluntary principles and guidelines could become laws

46 Artificial intelligence for science

7 Conclusion
This report has shown how AI application and development within computer
science, and all other major fields of science and research, has increased
substantially. The growth has been strongest over the past several years.
The coming decade is likely to see the growth continue, and AI become more
deeply and broadly adopted in most scientific research domains. References to AI
techniques in the titles, abstracts and keywords of research papers are likely to
(at some point) decline as the technology becomes commonplace. At this point
AI technology will be subsumed into application domains as ‘business as usual’.

The current surge in AI activity is not without historical The implications for science organisations arising from
precedent. Twice before in history AI research, investment this report are captured under the future development
and activity has surged. The peaks were followed by pathways of AI for science. Overall, science organisations
troughs; two AI winters are generally considered to have an imperative to upgrade AI capability to remain
have occurred during 1974–1980 and 1987–1993. Many of competitive and capable for the future. This will require
the conditions leading up to these two winters are education, training, hardware and software upgrades.
present today; however, there are significant differences. It will require the development of data assets and
The sheer size and momentum of the current AI boom is changed ways of working to become a more data-driven
unlikely to end anytime soon. There’s so much AI-related organisation. It will also require ensuring the development
investment, upskilling, organisational change and policy and application of AI is ethically sound and responds
development that it will be some time before it levels-off. to societal expectations, regulations and legislation.

The uptake of AI technologies within science and research The notion that AI will be doing research by itself
domains holds the promise of productivity improvement. seems unlikely. Scientific research requires creativity,
This is much needed as the global science sector is amid judgement, logic and communication skills that lie
an ongoing productivity slump where more research beyond the reach of current and foreseeable future
effort is being invested to achieve the same (or fewer) AI capability. However, human scientists working in
outcomes. The science productivity slump is causing a harmony with powerful AI technologies (where AI
broader productivity slump across most industries and the augments human capabilities), are likely to achieve better
entire economy. As a general-purpose technology, AI can outcomes, such as a higher rate of scientific discovery.
improve productivity in all domains of science and research
and, therefore, all industries. However, at this stage we still The economic depression of 1920–1921 resulting from
refer to this as ‘a promise’ for productivity uplift. There is the ‘Spanish Flu’ of 1918–1920 was followed by the
much evidence from case studies that AI is improving ‘Roaring 20s’, a decade of unprecedented economic
the efficiency and effectiveness of science, enabling growth. Economic historians [204] studying the
discoveries to happen faster, safer and at lower cost. Roaring 20s identify the general-purpose technology
However, this empirical evidence is currently not sufficient of electricity as the primary driver of productivity
as incontrovertible proof of the productivity gains of AI. growth in manufacturing. This productivity growth in
manufacturing stimulated overall economic growth
with spectacular results. It is possible that AI is the
general-purpose technology of our time which leads to
improved productivity in science which, in turn, improves
productivity and growth in the whole economy.

47
8 References
1. Crew, B. (2020) Google Scholar reveals its most 13. Boeing, P. and P. Hünermund (2020) A global
influential papers for 2020 - Artificial intelligence decline in research productivity? Evidence from
papers amass citations more than any other China and Germany. Economics Letters 197: 1-4.
research topic. Nature News (13 July). London.
14. Zhang, D., et al. (2021) The AI Index 2021 Annual
2. OECD (2022) OECD Statistics. Organisation for Report. AI Index Steering Committee, Human-Centered AI
Economic Cooperation and Development, Statistics Website Institute, Stanford University. Stanford CA, United States.
(https://stats.oecd.org/) accessed 11 March 2022. Paris.
15. Littman, M., et al. (2021) Gathering Strength,
3. OECD (2021) Database of national AI policies. Gathering Storms: The One Hundred Year Study on
Organisation for Economic Cooperation and Artificial Intelligence (AI100) Stanford University
Development, Artificial Intelligence Policy Observatory (http://ai100.stanford.edu/2021-report, website
(website accessed on 13/01/2022). Paris. accessed 4 March 2022). California, United States.

4. Van Roy, V., et al. (2021) AI Watch - National strategies 16. Nichols, J., et al. (2019) Artificial intelligence
on Artificial Intelligence: A European perspective, for science. Argonne National Laboratory.
2021 edition. European Comission. Publications Lemont, Illinois, United States.
Office of the European Union, Luxembourg.
17. Nicholson, K. and A. Slonim (2022) Special
5. Malcolm, E. (2021) Flexible solar panels: new stretch Report - Artificial intelligence: Your questions
of the imagination. CSIROscope (18 May 2021). Canberra. answered. The University of Adelaide Australian
Institute for Machine Learning, Australian
6. Bayat, A., et al. (2020) VariantSpark: Strategic Policy Research Institute. Adelaide.
Cloud-based machine learning for association
study of complex phenotype and large-scale 18. Gil, Y. and A. Selman (2019) A 20-Year Community
genomic data. GigaScience 9(8): 1-12. Roadmap for Artificial Intelligence Research in
the US. Computing Community Consortium (CCC)
7. Roberts, M., et al. (2021) Common pitfalls and and Association for the Advancement of Artificial
recommendations for using machine learning to detect Intelligence (AAAI). Washington DC, United States.
and prognosticate for COVID-19 using chest radiographs
and CT scans. Nature Machine Intelligence 3(3): 199-217. 19. Kaul, V., S. Enslin, and S.A. Gross (2020)
History of artificial intelligence in medicine.
8. Nicholson, K. and A. Slonim (2022) Artificial Gastrointestinal Endoscopy 92(4): 807-812.
intelligence: your questions answered. Australian
Strategic Policy Institute. Australia. 20. Haenlein, M. and A. Kaplan (2019) A Brief
History of Artificial Intelligence: On the Past,
9. UKRI (2022) Strategic Priorities Fund - Artificial Present, and Future of Artificial Intelligence.
intellgience - AI and Data Science for Science, Engineering, California Management Review 61(4): 5-14.
Health and Government. United Kingdom Research
and Innovation, United Kingdom Government (website 21. Buchanan, B.G. (2005) A (Very) Brief History
accessed 4 March 2022, https://www.ukri.org/our-work/ of Artificial Intelligence. AI Magazine 26(4): 53.
our-main-funds/strategic-priorities-fund/). London.
22. Flasiński, M. (2016) History of Artificial Intelligence,
10. AI4SD (2021) Artificial Intelligence and Augmented in Introduction to Artificial Intelligence, M. Flasiński,
Intelligence for Automated Investigations for Scientific Editor, Springer International Publishing: London.
Discovery (AI4SD). University of Southampton (https://www.
ai3sd.org, website accessed 7 March 2022). United Kingdom. 23. Wikipedia (2022) History of artificial
intelligence. Wikipedia - The free encyclopedia
11. Nolan, A. (2021) Artificial intelligence and (website accessed 22 Jan 2022). San Francisco.
the future of science. Organisation for Economic
Cooperation and Development, Artificial Intelligence 24. McCulloch, W.S. and W. Pitts (1943) A logical
Policy Observatory (25 October). Paris. calculus of the ideas immanent in nervous activity. The
bulletin of mathematical biophysics 5(4): 115-133.
12. Miyagawa, T. and T. Ishikawa (2019) On the Decline
of R&D Efficiency (RIETI Discussion Paper Series 19-E-052). 25. Minksy, M. and D. Edmonds (1954) Stochastic
Research Institute of Economy, Trade and Industry. Japan. Neural Analogue Reinforcement Calculator.
Princeton University. United States.

48 Artificial intelligence for science

26. Turing, A. (1950) Computing machinery 38. Jain, S. (2018) An Overview of Regularization
and intelligence. Mind - A quarterly review of Techniques in Deep Learning (with Python
psychology and philosophy 59(236): 433-460. code). Analytics Vidhya. Gurugram, India.

27. Warwick, K. and H. Shah (2016) Can machines 39. Zhang, Z. (2018) Improved Adam Optimizer
think? A report on Turing test experiments at for Deep Neural Networks. 2018 IEEE/ACM 26th
the Royal Society. Journal of Experimental & International Symposium on Quality of Service
Theoretical Artificial Intelligence 28(6): 989-1007. (IWQoS). DOI: 10.1109/IWQoS.2018.8624183.

28. McCarthy, J. (2006) The Dartmouth Workshop- 40. Mukkamala, M.C. and M. Hein, Variants of
-as planned and as it happened. Stanford University RMSProp and Adagrad with Logarithmic Regret
(website accessed 25 Jan 2022). United States. Bounds, in Proceedings of the 34th International
Conference on Machine Learning, P. Doina and
29. Agar, J.O.N. (2020) What is science for? The Lighthill T. Yee Whye, Editors. 2017, PMLR: Proceedings of
report on artificial intelligence reinterpreted. The British Machine Learning Research. p. 2545--2553.
Journal for the History of Science 53(3): 289-310.
41. Bottou, L. (2012) Stochastic Gradient Descent
30. Odagiri, H., Y. Nakamura, and M. Shibuya Tricks, in Neural Networks: Tricks of the Trade: Second
(1997) Research consortia as a vehicle for basic Edition, G. Montavon, G.B. Orr, and K.-R. Müller, Editors,
research: The case of a fifth generation computer Springer Berlin Heidelberg: Berlin, Heidelberg.
project in Japan. Research Policy 26(2): 191-207.
42. AAS (2022) What is science? (Reviewed by Pauline
31. Newquist, H. (2020) The Brain Makers: The History of Ladiges and Oliver Mayo). The Australian Academy of
Artificial Intelligence – Genius, Ego, And Greed In The Quest Science (website accessed 4 February 2022). Canberra.
For Machines That Think. The Relayer Group. New York.
43. UNESCO (2017) Measuring Scientific and Technological
32. lens.org (2022) Search, analyse and manage Services (STS): Draft Paper for Consultation. United
patent and scholarly data - Lens serves global Nations Educational, Scientific and Cultural Organisation,
patent and scholarly knowledge as a public good Institute for Statistics. Montreal, Quebec, Canada.
to inform science and technology enabled problem
solving. The Lens Website by Cambia (www.lens.org) 44. UIS (2022) The UNESCO Institute for
accessed on 8 March 2022. Brisbane, Australia. Statistics UIS.Stat. United Nations Economic
and Social Council website. Paris, France.
33. Jefferson, O.A., et al. (2021) Mapping CRISPR-
Cas9 public and commercial innovation using The Lens 45. PwC (2019) The 2018 global innovation
institutional toolkit. Transgenic Research 30(4): 585-599. 1000 study. PwC. London, UK.

34. Dutton, T., B. Barron, and G. Boskovic (2018) Building 46. Bajpai, P. (2021) Which companies spend the most in
an AI world. Report on national and regional AI strategies. research and development? NASDAQ News. New York City.
Canadian Institute for Advanced Research. Toronto, Canada.
47. Scopus (2021) What are Scopus subject area categories
35. Hajkowicz, S., et al. (2019) Artificial Intelligence: and ASJC codes? Elsevier Website (accessed on 3 March
Solving problems, growing the economy and improving 2022, last updated 29 July 2021). Amsterdam, Netherlands.
our quality of life. CSIRO. Brisbane, Australia.
48. Gao, J., et al. (2021) Potentially long-
36. IDC (2021) IDC Forecasts Companies to lasting effects of the pandemic on scientists.
Spend Almost $342 Billion on AI Solutions in 2021 Nature Communications 12(1): 6188.
(Media Release 4 August 2021). International Data
Corporation. Needham, MA, United States. 49. Riccaboni, M. and L. Verginer (2022) The impact
of the COVID-19 pandemic on scientific research
37. OECD (2021) OECD Science, Technology and in the life sciences. PLOS ONE 17(2): e0263001.
R&D Statistics: Revealed technology advantage
in selected fields. OECD Science, Technology 50. arXiv (2022) arXiv usage statistics. arXiv website
and R&D Statistics (database). Paris. (accessed 1 July 2022), Cornell University. United States.

49
51. PWC (2022) Trends - Paper Implementations 63. Kitano, H. (2021) Nobel Turing Challenge:
grouped by framework. Papers With Code Website creating the engine for scientific discovery. npj
(https://paperswithcode.com/trends). United States. Systems Biology and Applications 7(1): 29.

52. Schwartz, S.J. and B.L. Zamboanga (2009) 64. Gil, Y., R. King, and H. Kitano (2020) Posing an AI
The Peer-Review and Editorial System: Ways to Fix Scientist Grand Challenge: Artificial Intelligence Systems
Something That Might Be Broken. Perspectives Capable of Nobel-Quality Discoveries (Workshop
on Psychological Science 4(1): 54-61. Summary developed by Workshop Attendees). The
Alan Turing Institute (February 2020). London.
53. ABS (2017) The Australian National Census for 2016 -
Data accessed through the ABS table builder professional. 65. ATI (2022) The Turing AI scientist grand challenge -
Australian Bureau of Statistics (ABS). Canberra. Developing AI systems capable of making Nobel quality
scientific discoveries highly autonomously at a level
54. ABS (2021) Research and experimental comparable, and possibly superior, to the best human
development, businesses, Australia. Australian scientists by 2050. The Alan Turing Institute, Research
Bureau of Statistics. Canberra, Australia. Projects (website accessed 6 April 2022). London.
55. Parham, D. (2009) Empirical Analysis of the 66. Kelling, S., et al. (2009) Data-intensive Science: A New
Effects of R&D on Productivity - Implications for Paradigm for Biodiversity Studies. BioScience 59(7): 613-620.
productivity measurement. Organisation for
Economic Cooperation and Development. Paris. 67. Leonelli, S. (2012) Classificatory Theory
in Data-intensive Science: The Case of Open
56. Bloom, N., et al. (2020) Are Ideas Getting Harder Biomedical Ontologies. International Studies
to Find? American Economic Review 110(4): 1104-44. in the Philosophy of Science 26(1): 47-65.
57. Towse, A., et al. (2017) Time for a change in how new 68. Ahrens, J., et al. (2011) Data-Intensive Science
antibiotics are reimbursed: Development of an insurance in the US DOE: Case Studies and Future Challenges.
framework for funding new antibiotics based on a policy Computing in Science & Engineering 13(6): 14-24.
of risk mitigation. Health Policy 121(10): 1025-1030.
69. Pietsch, W. (2015) Aspects of Theory-Ladenness in
58. Kraus, C.N. (2008) Low hanging fruit in Data-Intensive Science. Philosophy of Science 82(5): 905-916.
infectious disease drug development. Current
Opinion in Microbiology 11(5): 434-438. 70. Nikam, V.B. and B.B. Meshram (2013) Modeling
Rainfall Prediction Using Data Mining Method: A
59. ITER (2022) Advantages of fusion - The next Bayesian Approach. Fifth International Conference
decades are crucially important to putting the world on Computational Intelligence, Modelling and
on a path of reduced greenhouse gas emissions. Simulation. DOI: 10.1109/CIMSim.2013.29.
International Thermonuclear Experimental Reactor
(website accessed 6 April 2022). France. 71. Hassani, H. and E.S. Silva (2015) Forecasting with
Big Data: A Review. Annals of Data Science 2(1): 5-19.
60. CCFE (2022) Fusion energy record demonstrates
powerplant future (9 February). Culham Centre for Fusion 72. Callebaut, W. (2012) Scientific perspectivism: A
Energy (CCFE), United Kingdom Energy Authority (website philosopher of science’s response to the challenge of
accessed 6 April 2022). Oxfordshire, United Kingdom. big data biology. Studies in History and Philosophy of
Science Part C: Studies in History and Philosophy of
61. Degrave, J., et al. (2022) Magnetic control of Biological and Biomedical Sciences 43(1): 69-80.
tokamak plasmas through deep reinforcement
learning. Nature 602(7897): 414-419. 73. Cowls, J. and R. Schroeder (2015) Causation,
Correlation, and Big Data in Social Science
62. Katwala, A. (2022) DeepMind Has Trained an AI to Research. Policy & Internet 7(4): 447-472.
Control Nuclear Fusion - The Google-backed firm taught
a reinforcement learning algorithm to control the fiery 74. Elragal, A. and R. Klischewski (2017) Theory-driven
plasma inside a tokamak nuclear fusion reactor. Wired or process-driven prediction? Epistemological challenges
Magazine (16 Feb 2022). San Francisco, United States. of big data analytics. Journal of Big Data 4(1): 19.

50 Artificial intelligence for science

75. Senior, A.W., et al. (2020) Improved protein 89. Castelvecchi, D. (2021) DeepMind’s AI helps untangle
structure prediction using potentials from the mathematics of knots. Nature 600(7888): 202-202.
deep learning. Nature 577(7792): 706-710.
90. Davies, A., et al. (2021) Advancing mathematics by
76. DeepMind (2020) AlphaFold: Using AI for guiding human intuition with AI. Nature 600(7887): 70-74.
scientific discovery. Deepmind (https://deepmind.
com) website accessed 27 April 2022. London. 91. Pascal, L., et al. (2020) A Shiny r app to solve
the problem of when to stop managing or surveying
77. Jumper, J., et al. (2021) Highly accurate protein structure species under imperfect detection. Methods
prediction with AlphaFold. Nature 596(7873): 583-589. in Ecology and Evolution 11(12): 1707-1715.

78. AlphaFold and EMBL-EBI, AlphaFold Protein 92. Torney, C.J., et al. (2019) A comparison of
Structure Database, A.a. EMBL-EBI, Editor. deep learning and citizen science techniques for
2022: Cambridgeshire, United Kingdom. counting wildlife in aerial survey images. Methods
in Ecology and Evolution 10(6): 779-787.
79. iNaturalist (2021) A Community for Naturalists.
iNaturalist Web page accessed 24 January 2022. Australia. 93. Norouzzadeh, M.S., et al. (2018) Automatically
identifying, counting, and describing wild animals in
80. Van Horn, G., et al. (2018) The inaturalist species camera-trap images with deep learning. Proceedings of
classification and detection dataset. Proceedings of the IEEE the National Academy of Sciences 115(25): E5716-E5725.
conference on computer vision and pattern recognition.
94. Wilhite, A.W. and E.A. Fong (2012) Coercive Citation
81. Ceccaroni, L., et al. (2019) Opportunities and risks in Academic Publishing. Science 335(6068): 542-543.
for citizen science in the age of artificial intelligence.
Citizen Science: Theory and Practice 4(1). 95. Seglen, P.O. (1997) Why the impact factor of
journals should not be used for evaluating research.
82. McClure, E.C., et al. (2020) Artificial Intelligence Meets BMJ (Clinical research ed.) 314(7079): 498-502.
Citizen Science to Supercharge Ecological Monitoring.
Patterns 1(7): https://doi.org/10.1016/j.patter.2020.100109. 96. Weis, J.W. and J.M. Jacobson (2021) Learning on
knowledge graph dynamics provides an early warning of
83. Open Science Collaboration (2015) Estimating impactful research. Nature Biotechnology 39(10): 1300-1307.
the reproducibility of psychological science.
Science 349(6251): doi:10.1126/science.aac4716. 97. Savage, N. (2019) How AI and neuroscience drive
each other forwards. Nature 571(7766): S15-S15.
84. Camerer, C.F., et al. (2016) Evaluating
replicability of laboratory experiments in 98. Yamins, D.L.K., et al. (2014) Performance-
economics. Science 351(6280): 1433-1436. optimized hierarchical models predict neural
responses in higher visual cortex. Proceedings of the
85. Prinz, F., T. Schlange, and K. Asadullah National Academy of Sciences 111(23): 8619-8624.
(2011) Believe it or not: how much can we rely
on published data on potential drug targets? 99. Kell, A.J.E., et al. (2018) A Task-Optimized Neural
Nature Reviews Drug Discovery 10(9): 712-712. Network Replicates Human Auditory Behavior,
Predicts Brain Responses, and Reveals a Cortical
86. Yang, Y., W. Youyou, and B. Uzzi (2020) Estimating Processing Hierarchy. Neuron 98(3): 630-644.
the deep replicability of scientific findings using
human and artificial intelligence. Proceedings of the 100. Rajpurkar, P., et al. (2022) AI in health and
National Academy of Sciences 117(20): 10762-10768. medicine. Nature Medicine (28): 31–38.

87. Stach, E., et al. (2021) Autonomous experimentation 101. Stokes, J.M., et al. (2020) A Deep Learning Approach
systems for materials development: A community to Antibiotic Discovery. Cell 180(4): 688-702.
perspective. Matter 4(9): 2702-2726.
102. RAR (2014) Antimicrobial Resistance:
88. Nikolaev, P., et al. (2016) Autonomy in materials Tackling a Crisis for the Health and Wealth of
research: a case study in carbon nanotube growth. Nations. Review on Antimicrobial Resistance,
npj Computational Materials 2(1): 16031. United Kingdom Government. London.

51
103. Brown, E.D. and G.D. Wright (2016) Antibacterial drug 116. Ellegaard, O. and J.A. Wallin (2015) The bibliometric
discovery in the resistance era. Nature 529(7586): 336-343. analysis of scholarly production: How great is the
impact? Scientometrics 105(3): 1809-1831.
104. Zhavoronkov, A., et al. (2019) Deep learning
enables rapid identification of potent DDR1 kinase 117. Hu, Q., S. Khosa, and N. Kapucu (2015)
inhibitors. Nature Biotechnology 37(9): 1038-1040. The Intellectual Structure of Empirical Network
Research in Public Administration. Journal of Public
105. VoPham, T., et al. (2018) Emerging trends Administration Research and Theory 26(4): 593-612.
in geospatial artificial intelligence (geoAI):
potential applications for environmental 118. Donthu, N., et al. (2021) How to conduct a
epidemiology. Environmental Health 17(1): 40. bibliometric analysis: An overview and guidelines.
Journal of Business Research 133: 285-296.
106. Lin, Y., et al. (2017) Mining public datasets for modeling
intra-city PM2. 5 concentrations at a fine spatial resolution. 119. Wang, K., et al. (2020) Microsoft Academic
Proceedings of the 25th ACM SIGSPATIAL international Graph: When experts are not enough.
conference on advances in geographic information systems. Quantitative Science Studies 1(1): 396-413.

107. Checco, A., et al. (2021) AI-assisted peer review. 120. Frank, M.R., et al. (2019) The evolution of
Humanities and Social Sciences Communications 8(1): 25. citation graphs in artificial intelligence research.
Nature Machine Intelligence 1(2): 79-85.
108. Frontiers (2020) Artificial Intelligence to help meet
global demand for high-quality, objective peer-review in 121. Liu, N., P. Shapira, and X. Yue (2021)
publishing. Frontiers Science News. Lausanne, Switzerland. Tracking developments in artificial intelligence
research: constructing and applying a new search
109. Vincent, J. (2022) A new use for AI: Summarizing strategy. Scientometrics 126(4): 3153-3192.
scientific research for seven-year-olds. The Verge (18 January
2022). Washington, District of Columbia, United States. 122. Baruffaldi, S., et al. (2020) Identifying and measuring
developments in artificial intelligence. Organisation
110. Bickler, S.H. (2021) Machine Learning for Economic Cooperation and Development. Paris.
Arrives in Archaeology. Advances in
Archaeological Practice 9(2): 186-191. 123. Penfold, R. (2020) Using the Lens Database
for Staff Publications. Journal of the Medical
111. Chetouani, A., et al. (2020) Classification of engraved Library Association : JMLA 108(2): 341-344.
pottery sherds mixing deep-learning features by compact
bilinear pooling. Pattern Recognition Letters 131: 1-7. 124. Jefferson, O.A., et al. (2021) Mapping
innovation trajectories on SARS-CoV-2 and its
112. Romanengo, C., S. Biasotti, and B. Falcidieno variants. Nature Biotechnology 39(4): 401-403.
(2020) Recognising decorations in archaeological finds
through the analysis of characteristic curves on 3D 125. Geroski, P.A. (2000) Models of technology
models. Pattern Recognition Letters 131: 405-412. diffusion. Research Policy 29(4): 603-625.

113. Bonhage, A., et al. (2021) A modified Mask region- 126. Reuther, A., et al. (2020) Survey of Machine
based convolutional neural network approach for Learning Accelerators. 2020 IEEE High Performance
the automated detection of archaeological sites Extreme Computing Conference (HPEC). DOI:
on high-resolution light detection and ranging- 10.1109/HPEC43674.2020.9286149.
derived digital elevation models in the North German
Lowland. Archaeological Prospection 28(2): 177-186. 127. Arute, F., et al. (2019) Quantum supremacy
using a programmable superconducting
114. Davis, D.S., R.J. DiNapoli, and K. Douglass (2020) processor. Nature 574(7779): 505-510.
Integrating Point Process Models, Evolutionary Ecology and
Traditional Knowledge Improves Landscape Archaeology—A 128. Korot, E., et al. (2021) Code-free deep learning
Case from Southwest Madagascar. Geosciences 10(287): 1-25. for multi-modality medical image classification.
Nature Machine Intelligence 3(4): 288-298.
115. Mirhoseini, A., et al. (2021) A graph placement
methodology for fast chip design. Nature 594(7862): 207-212. 129. Microsoft (2022) Artificial Intelligence
sample for Power BI: Take a tour (2 April 2022).
Microsoft Corporation. Seattle, United States.

52 Artificial intelligence for science

130. Microsoft (2018) Bringing AI to Excel—4 new Review (18 October 2016). Massachusetts
features announced today at Ignite. Microsoft Institute of Technology, Cambridge, MA.
Corporation. Seattle, United States.
144. Anderson, C. (2015) Creating a data-driven
131. Martin, M. (2022) Fifteen best Github organisation. O’Rielly Media Inc. Sebastopol, California.
alternatives in 2022. Guru99 Blog. Wilmington,
Delaware, United States of America. 145. OECD (2022) OECD.AI Policy Observatory:
AI courses in English by educational level. OECD.
132. Bojer, C.S. and J.P. Meldgaard (2021) Kaggle forecasting AI Policy Observatory. Paris, France.
competitions: An overlooked learning opportunity.
International Journal of Forecasting 37(2): 587-603. 146. Nature Index (2020) Top 100 academic institutions
in artificial intelligence. Nature Index 2020. London.
133. Mishkin, D., N. Sergievskiy, and J. Matas
(2017) Systematic evaluation of convolution neural 147. University of Adelaide (2020) Australian
network advances on the Imagenet. Computer Institute for Machine Learning. Annual report
Vision and Image Understanding 161: 11-19. 2020. University of Adelaide. Adelaide.

134. Gartner (2021) Gartner Forecasts Worldwide 148. DandoloPartners (2020) Evaluation of early
Public Cloud End-User Spending to Grow 23% in learning and schools initiatives in the National
2021. Gartner Newsroom - Press Release 21 April Innovation and Science Agenda. Report to
2021. Stamford, Connecticut, United States. the Department of Education. Department of
Education, Australian Government. Canberra.
135. ReportLinker (2021) Cloud Computing Market
by Service, Deployment Model, Organization 149. Toney, A. and M. Flagg (2020) U.S. Demand
Size, Vertical And Region - Global Forecast to for AI-Related Talent. Center for Security and
2026. ReportLinker- Market Ingelligence Platform Emerging Technology (CSET). Washington D.C.
(https://ai.reportlinker.com/) Lyon, France. 150. Mason, C., et al. (2022) Australian Skills
136. Drake, N. (2014) Cloud computing beckons Dashboard. Commonwealth Scientific and
scientists. Nature 509(7502): 543-544. Industrial Research Organisation. Brisbane.

137. Langmead, B. and A. Nellore (2018) Cloud 151. OECD (2022) Live data: Cross-country AI
computing for genomic data analysis and collaboration. skills penetration (accessed on 27/01/2022).
Nature Reviews Genetics 19(4): 208-219. Organisation for Economic Cooperation and
Development, AI Policy Observatory. Paris.
138. Wynants, L., et al. (2020) Prediction models
for diagnosis and prognosis of covid-19: systematic 152. MacroPolo (2022) The Global AI Talent Tracker.
review and critical appraisal. 369: m1328. MacroPolo (https://macropolo.org/digital-projects/
the-global-ai-talent-tracker). Chicago.
139. Floridi, L. (2019) What the Near Future of Artificial
Intelligence Could Be. Philosophy & Technology 32(1): 1-15. 153. DESE (2022) Award Course Completions
Time Series. Department of Education, Skills and
140. De Fauw, J., et al. (2018) Clinically applicable Employment, Australian Government. Canberra.
deep learning for diagnosis and referral in retinal
disease. Nature Medicine 24(9): 1342-1350. 154. Hajkowicz, S., et al. (2016) Tomorrow’s Digitally
Enabled Workforce: Megatrends and scenarios for
141. Ramarao-Milne, P., et al. (2022) Data-driven jobs and employment in Australia over the coming
platform for identifying variants of interest in twenty years. CSIRO. Brisbane, Australia.
COVID-19 virus. Computational and Structural
Biotechnology Journal 20: 2942-2950. 155. Zhang, B. and A. Dafoe, Artificial intelligence:
American attitudes and trends. 2019, University
142. Bauer, D.C., et al. (2021) Interoperable medical of Oxford: Oxford, United Kingdom.
data: The missing link for understanding COVID-19.
Transboundary and Emerging Diseases 68(4): 1753-1760. 156. Gillespie, N., S. Lockey, and C. Curtis (2021) Trust in
artificial Intelligence: a five country study. University
143. Kiron, D. (2016) Lessons from Becoming a of Queensland and KPMG. Brisbane, Australia.
Data-Driven Organization. MIT Sloan Management

53
157. Kusters, R., et al. (2020) Interdisciplinary 169. Wang, D., et al. (2019) Human-AI collaboration
Research in Artificial Intelligence: Challenges in data science: Exploring data scientists’ perceptions
and Opportunities. Frontiers in Big Data 3: DOI: of automated AI. Proceedings of the ACM on
https://doi.org/10.3389/fdata.2020.577974. Human-Computer Interaction 3(CSCW): 1-24.

158. Wang, M.-T. and J.L. Degol (2017) Gender 170. Procter, R., B. Glover, and E. Jones (2020)
Gap in Science, Technology, Engineering, and Research 4.0: Research in the age of automation.
Mathematics (STEM): Current Knowledge, Implications Demos. London, United Kingdom.
for Practice, Policy, and Future Directions.
Educational Psychology Review 29(1): 119-140. 171. Chubb, J., P. Cowling, and D. Reed (2021) Speeding up to
keep up: exploring the use of AI in the research process. AI
159. Simoncini, K. and M. Lasen (2018) Ideas About & Society DOI: https://doi.org/10.1007/s00146-021-01259-0.
STEM Among Australian Early Childhood Professionals:
How Important is STEM in Early Childhood Education? 172. Feuston, J.L. and J.R. Brubaker (2021) Putting Tools in
International Journal of Early Childhood 50(3): 353-369. Their Place: The Role of Time and Perspective in Human-AI
Collaboration for Qualitative Analysis. Proceedings of the
160. Tai, R.H., et al. (2006) Planning early for ACM on Human Computer Interaction 5(CSCW2): 1-25.
careers in science. Science 312(5777): 1143-1144.
173. Jiang, J.A., et al. (2021) Supporting Serendipity:
161. Maltese, A.V. and R.H. Tai (2011) Pipeline Opportunities and Challenges for Human-AI Collaboration
persistence: Examining the association of educational in Qualitative Analysis. Proceedings of the ACM on
experiences with earned degrees in STEM among Human-Computer Interaction 5(CSCW1): 1-23.
US students. Science education 95(5): 877-907.
174. Hervieux, S. and A. Wheatley (2021) Perceptions
162. McKinsey & Company (2017) Jobs lost, jobs of artificial intelligence: A survey of academic
gained: Workforce transitions in a time of automation. librarians in Canada and the United States. The
McKinsey & Company. New York, United States. Journal of Academic Librarianship 47(1): 1-11.

163. Frey, C.B. and M.A. Osborne (2017) The future of 175. Lund, B.D., et al. (2020) Perceptions toward Artificial
employment: how susceptible are jobs to computerisation? Intelligence among Academic Library Employees
Technological Forecasting and Social Change 114(1): 254-280. and Alignment with the Diffusion of Innovations’
Adopter Categories. College & Research Libraries
164. Dellermann, D., et al. (2019) The Future of 81(5): DOI: https://doi.org/10.5860/crl.81.5.865.
Human-AI Collaboration: A Taxonomy of Design
Knowledge for Hybrid Intelligence Systems. 176. Wood, B.A. and D. Evans (2018) Librarians’ perceptions
arXiv (arXiv:2105.03354). Cornell University. of artificial intelligence and its potential impact on the
profession. Kennesaw State University. Kennesaw, GA.
165. Shook, E. and M. Knickrehm (2018) Reworking
the revolution. Accenture. Sydney. 177. Cox, A.M., S. Pinfield, and S. Rutter (2019) The
intelligent library: Thought leaders’ views on the
166. Partnership on AI (2019) Collaborations likely impact of artificial intelligence on academic
Between People and AI Systems (CPAIS): Human libraries. Library Hi Tech 37(3): 418-435.
- AI Collaboration Framework and Case Studies.
Partnership on AI (https://partnershiponai.org/). 178. Selwyn, N., et al. (2020) AI for Social Good-Australian
San Francisco, California, United States. Attitudes Toward AI and Society Report. pdf. Monash Data
Futures Institute, Monash University. Victoria, Australia.
167. Castagno, S. and M. Khalifa (2020) Perceptions
of Artificial Intelligence Among Healthcare Staff: A 179. Oh, S., et al. (2019) Physician Confidence in Artificial
Qualitative Survey Study. Frontiers in Artificial Intelligence Intelligence: An Online Mobile Survey. Journal of
3(84): DOI: https://doi.org/10.3389/frai.2020.578983. Medical Internet Research 21(3): DOI: 10.2196/12422.

168. European Society of Radiology (2019) Impact of 180. Jarrahi, M.H. (2018) Artificial intelligence and the
artificial intelligence on radiology: a EuroAIM survey among future of work: Human-AI symbiosis in organizational
members of the European Society of Radiology. Insights decision making. Business Horizons 61(4): 577-586.
Imaging 10(105): https://doi.org/10.1186/s13244-019-0798-3.

54 Artificial intelligence for science

181. Lubars, B. and C. Tan (2019) Ask not what AI can 194. Jobin, A., M. Ienca, and E. Vayena (2019)
do, but what AI should do: Towards a framework of The global landscape of AI ethics guidelines.
task delegability. arXiv preprint arXiv:1902.03245. Nature Machine Intelligence 1(9): 389-399.

182. Krittanawong, C. (2018) The rise of artificial 195. DISER (2022) Australia’s Artificial Intelligence
intelligence and the uncertain future for Ethics Framework. Department of Industry, Science and
physicians. Eur J Intern Med 48: e13-e14. Energy Resources, Australian Government. Canberra.

183. Drozdal, J., et al. (2020) Trust in AutoML: 196. Sanderson, C., et al. (2022) AI Ethics
Exploring Information Needs for Establishing Trust Principles in Practice: Perspectives of Designers
in Automated Machine Learning Systems. and Developers. arXiv:2112.07467v2 (https://
doi.org/10.48550/arXiv.2112.07467).
184. Stathoulopoulos, K. and J. Mateos-Garcia (2019)
Gender Diversity in AI Research. National Endowment 197. Burt, A. (2021) New AI Regulations Are
for Science, Technology and the Arts. London. Coming. Is Your Organization Ready? Harvard
Business Review (April 30, 2021).
185. DISER (2021) STEM Equity Monitor 2021, a national
data report on girls’ and women’s participation in 198. Treasury (2021) Request for Information and Comment
science, technology, engineering and mathematics on Financial Institutions’ Use of Artificial Intelligence,
(STEM). Australian Government Department of Including Machine Learning. Department of the Treasury,
Industry Science Energy and Resources. Canberra. United States Government. Washington DC, United States.

186. STA (2020) Diversity gains in Australia’s STEM 199. EC (2021) Proposal for a Regulation laying
workforce - But more needed. Science and Technology down harmonised rules on artificial intelligence.
Australia (website accessed 8 February 2022). Canberra. European Commission (Policy and Legislation,
Publication 21 April 2021). Brussels.
187. Ball, R. (2015) STEM the gap:Science belongs to us
mob too. Australian Quarterly January-March: 13-19. 200. Angelov, P.P., et al. (2021) Explainable artificial
intelligence: an analytical review. 11(5): e1424.
188. Latimer, J., et al. (2019) Australia’s strategy to achieve
gender equality in STEM. The Lancet 393(10171): 524-526. 201. Briggs, C., Z. Fan, and P. Andras (2021) A Review of
Privacy-Preserving Federated Learning for the Internet-
189. CSIRO (2022) Indigenous science (web page). CSIRO of-Things, in Federated Learning Systems: Towards
(Website accessed 8 February 2022, https://www.csiro. Next-Generation AI, M.H. Rehman and M.M. Gaber,
au/en/research/indigenous-science). Canberra. Editors, Springer International Publishing: London.
190. CSIRO (2019) AI transforms Kakadu management 202. Boulemtafes, A., A. Derhab, and Y. Challal
- CSIRO, Kakadu rangers and Microsoft meld science, (2020) A review of privacy-preserving techniques
Indigenous knowledge and technology in pioneering for deep learning. Neurocomputing 384: 21-45.
program. CSIRO Media Release (20 November). Canberra.
203. Fahse, T., V. Huber, and B. van Giffen (2021)
191. Schiff, D., et al. (2021) AI Ethics in the Managing Bias in Machine Learning Projects.
Public, Private, and NGO Sectors: A Review of a Innovation Through Information Systems.
Global Document Collection. IEEE Transactions Cham. Springer International Publishing.
on Technology and Society 2(1): 31-42.
204. Harrison, S. and M. Weder (2009) Technological
192. Schiff, D., et al. (2020) What’s Next for AI change and the roaring twenties: A neoclassical
Ethics, Policy, and Governance? A Global Overview. perspective. Journal of Macroeconomics 31(3): 363-375.
Proceedings of the AAAI/ACM Conference on AI, Ethics,
and Society. New York. Association for Computing
Machinery. DOI: 10.1145/3375627.3375804.

193. Hagendorff, T. (2020) The Ethics of AI Ethics: An

Evaluation of Guidelines. Minds and Machines 30(1): 99-120.

55
Appendix A –
Science occupations
Considering the four types of science (natural science, The category of natural and physical science
social science, formal science and applied science) we professionals contains sub-categories of
identified corresponding occupations from the Australian
and New Zealand Standard Classification of Occupations • agricultural, fisheries and forestry scientists (2341)
(ANZSCO). We sought occupations where the description • chemists, and food and wine scientists (2342)
involved research activity and was consistent with tasks • environmental scientists (2343)
comprising the scientific process. The occupations we • geologists, geophysicists and hydrogeologists (2344)
identified (with ANZSCO codes in brackets) included:
• life scientists (2345)
• science technicians (3114) • medical scientists (2346)
• natural and physical science professionals (234) • veterinarians (2347)
• economists (224311) • and other natural and physical
• social professionals (272499) science professionals (2349).
• mathematicians (224112) The ‘social professional’ category captures social scientists
• statisticians (224113). conducting research in diverse fields: anthropologists,
criminologists, geographers, political scientists,
sociologists and others. We note that some of these
categories may contain relatively small numbers of
non‑science occupations that we could not separate out.

Some of our talented people who work on robotics and autonomous systems.

56 Artificial intelligence for science

Appendix B –
Artificial intelligence phrases
During 2018–2019 the Organisation for Economic Co-operation and Development (OECD) held several expert working
groups to identify, review and refine a set of phrases to search for artificial intelligence (AI) patents and scholarly
publications. In the final meeting in 2019 the patent examiners and AI experts convened by the OECD went through
the final draft list of phrases to validate and challenge them. The final set of phrases was published by the OECD [122].
We have also used these phrases to search for AI publications in this report. The phrases are listed below.

action recognition cognitive computing feature engineering

activity recognition cognitive insight system feature extraction

adaboost cognitive modelling feature learning

adaptive boosting collaborative filtering feature selection

adversarial network collision avoidance firefly algorithm

ambient intelligence community detection fuzzy c

ant colony computational intelligence fuzzy environment

ant colony optimisation computational pathology fuzzy logic

artificial bee colony algorithm computer vision fuzzy number

artificial intelligence convolutional neural network fuzzy set

artificial neural network cyber physical system fuzzy system

association rule data mining gaussian mixture model

autoencoder decision tree gaussian process

autonomic computing deep belief network generative adversarial network

autonomous vehicle deep convolutional neural network genetic algorithm

autonomous weapon deep learning genetic programming

backpropagation deep neural network gesture recognition

Bayesian learning dictionary learning gradient boosting

bayesian network differential evolution algorithm gradient tree boosting

bee colony dimensionality reduction graphical model

biped robot dynamic time warping gravitational search algorithm

blind signal separation emotion recognition hebbian learning

bootstrap aggregation ensemble learning hidden Markov model

brain computer interface evolutionary algorithm hierarchical clustering

brownboost evolutionary computation high-dimensional data

chatbot extreme machine learning high-dimensional feature

classification tree face recognition high-dimensional input

cluster analysis facial expression recognition high-dimensional model

cognitive automation factorisation machine high-dimensional space

57
high-dimensional system lpboost particle swarm optimisation

human action recognition machine intelligence pattern recognition

human activity recognition machine learning pedestrian detection

human aware artificial intelligence machine translation policy gradient methods

humanoid robot machine vision Q-learning

human-robot interaction madaboost quadruped robot

image classification MapReduce random field

image processing Markovian random forest

image recognition memetic algorithm rankboost

image retrieval meta learning recommender system

image segmentation motion planning recurrent neural network

independent component analysis multi task learning regression tree

inductive monitoring multi-agent system reinforcement learning

industrial robot multi-label classification relational learning

instance-based learning multi-layer perceptron robot

intelligence augmentation multinomial naive Bayes rough set

intelligent agent multi-objective evolutionary algorithm rule learning

intelligent classifier multi-objective optimisation rule-based learning

intelligent geometric computing multi-sensor fusion self-organising map

intelligent infrastructure naive Bayes classifier self-organising structure

intelligent software agent natural gradient semantic web

intuitionistic fuzzy set natural language generation semi-supervised learning

Kernel learning natural language processing sensor data fusion

K-means natural language understanding sensor fusion

latent dirichlet allocation nearest neighbour algorithm sentiment analysis

latent semantic analysis neural network service robot

latent variable neural turing similarity learning

layered control system neural turing machine simultaneous localisation mapping

learning automata neuromorphic computing single-linkage clustering

legged robot non negative matrix factorisation social robot

link prediction object detection sparse representation

logitboost object recognition spectral clustering

long short term memory (LSTM) obstacle avoidance speech recognition

58 Artificial intelligence for science

speech to text

stacked generalisation

statistical relational learning

stochastic gradient

supervised learning

support vector machine

support vector regression

swarm intelligence

swarm optimisation

t s fuzzy system

Takagi-Sugeno fuzzy systems

temporal difference learning

text mining

text to speech

topic model

totalboost

trajectory planning

trajectory tracking

transfer learning

trust region policy optimisation

unmanned aerial vehicle

unsupervised learning

variational inference

vector machine

virtual assistant

visual servoing

wheeled mobile robot

xgboost

59
As Australia’s national science
agency and innovation catalyst,
CSIRO is solving the greatest
challenges through innovative
science and technology.
CSIRO. Unlocking a better future
for everyone.

Contact us
1300 363 400
+61 3 9545 2176
csiro.au/contact
csiro.au

For further information

Stefan Hajkowicz
Principal Scientist in Strategy and Foresight
[email protected]
csiro.au/data61

B&M | 22-00612