Foresight 4. Artificial Intelligence

Background

Progress in Artificial Intelligence (AI) research has accelerated in the last 15 years, with notable achievements including robust dictation and translation software, IBM’s Watson winning Jeopardy, driverless cars, and the defeat of the world Go champion by a deep-learning algorithm. This has in turn prompted an avalanche of popular books, articles and reports assessing the likely impact of future advances in AI and robotics on human jobs. Many quote a 2013 University of Oxford study that concluded 47% of US jobs are at risk of being eliminated by automation in the next 20 years. The spectrum of jobs thought to be vulnerable to automation has expanded from menial, low-paid, repetitive jobs to some highly-skilled and remunerated occupations: e.g. Google-type search engines replace para-legal staff with law degrees. Susskind and Susskind (2015) concluded that the professions in particular are highly vulnerable to technological disruption.

Scientists might take comfort from the fact that, of all the professions, science relies most on human intelligence and creativity, the very attributes that are regarded as most difficult to achieve in AI. But experts in Go describe moves by Deepmind’s AI as showing brilliant innovation and intuition. Domingos (2015) argues that machine learning algorithms are already capable of analysing big data sets, searching knowledge databases, developing complex models and designing experimental protocols, in ways that are beyond the abilities of individual humans.

We might take comfort from the observation that science has been continually disrupted by advances in technology for hundreds of years, and that the feedback loop from science to technology and back to science is a key reason for its success. The next few decades may just seem like an accelerated version of that process. The interesting question is whether the adoption of AI will lead to a qualitative shift in the institutions, practice and productivity of science.

Scenario

Over the next 20 years, competition among scientists, research agencies and research users will drive the rapid uptake and adoption of AI and automation within science. This process is already underway, and many of the following scenario elements represent extensions of trends that are already evident.

• Observational science becomes dominated by autonomous sensors and instruments (like the Argo ocean profiling floats). As an AI-directed Internet of Things grows, the distinction between “scientific” and “civilian” observations becomes blurred, especially in environmental and health sciences.
• Laboratory and field experiments are increasingly conducted by robotic systems (e.g. automated DNA analysis and synthesis), using protocols and testing hypotheses produced by AI algorithms.
• Scientific modelling is dominated by large, “smart”, data-assimilating and self-defining community models, and/or automated, user-friendly, accessible model packages.
• There is increased adoption of open data policies, and intelligent and flexible data analysis and modelling tools are available online to all, including non-scientists, leading to a rise in “citizen science”.
• The scientific literature is open and online, and science search engines become more sophisticated and intelligent, able to provide useful and interpretive answers to broad enquiries by users. Scientific papers are generated by AI, initially with human “co-authors”.
• Traditional science providers increasingly compete with online providers of research services, who rely on AI-directed automated laboratories, public data and remote deployable field sensors to deliver effective and efficient solutions. Research users bypass normal funding channels and research brokers, and obtain solutions directly from online providers.

Indicators: How would we know this was “starting to happen”?

If one accepts the argument that science will inevitably continue to be disrupted by new technologies, including AI and automation, then we are not looking for indicators that this is “starting to happen”, but for indicators suggesting the pace is accelerating.

1. The proportion of scientific data obtained from autonomous, intelligent sensors and robots constitutes more than 80% of marine observations.
2. Scientific hypotheses and experimental designs are generated by smart algorithms in 50% or more of scientific papers.
3. Fifty percent or more of model applications rely on scientific models which are self-defining and refining, data-assimilating, widely accessible online and user friendly.
4. Twenty percent or more of scientific papers are generated by AI, with little or no human input.
5. Online providers of AI-directed, largely automated research services are responsible for 50% or more of applied scientific research.

Scoring of indicators

Project team only – “score” this scenario (requires login): Click to continue.

Additional Reading

Brynjolfsson, E and A. McAfee. 2014. The Second Machine Age. Work, Progress and Prosperity in a Time of Brilliant Technology.  Norton & Co., NY.

Domingos, P. 2015. The Master Algorithm. 322pp. Penguin Press.

Osborne, M.A. and C.B. Frey. 2013. The Future of Employment: How Susceptible are Jobs to Computerisation. http://www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf

Susskind, R. and D. Susskind. 2015. The Future of the Professions. How Technology will Transform the Work of Human Experts. Oxford University Press.

Examples of AI in science

June 2018 – Using AI to cope with increased volume of scientific publications – IPCC example