Science in the age of Artificial Intelligence
Artificial intelligence holds massive promise for human advancement but it’s not infallible and needs ethics, regulation and governance to ensure it is resilient and not manipulated. “We need to apply science to understand the properties and impact of emerging AI systems. Discussions must also involve a plurality of perspectives if we are to realise the full potential of AI.” This is the opinion of STIAS Donald Gordon Fellow Professor Sir Nigel Shadbolt who was presenting the first STIAS public lecture of 2024.
Shadbolt is a world-leading researcher in Artificial Intelligence (AI). He has been working in AI for over four decades, having obtained his PhD from the University of Edinburgh’s Department of Artificial Intelligence in 1984. He is Principal of Jesus College Oxford, Professor of Computing Science at the University of Oxford and chairman of the Open Data Institute which he co-founded with Sir Tim Berners-Lee. In 2009 he was appointed Information Advisor by the UK Prime Minister and, with Berners-Lee, oversaw Open Data releases in the public sector. He was knighted in 2013 for ‘services to science and engineering’.
At Oxford he has focused his research on human-centred AI in various applications. Most recently he was asked to lead the setting up of the Oxford Institute of Ethics in AI. With over 500 publications, he has researched and published on topics ranging from cognitive psychology to computational neuroscience, AI to the Semantic Web. In 2018 he published The Digital Ape: how to live (in peace) with smart machines, described as a ‘landmark book’. He is a Fellow of the Royal Society, the Royal Academy of Engineering, and the British Computer Society.
His wide-ranging lecture reflected on what makes the scientific method so powerful, how it has transformed our world, and how it has been informed, enriched and made possible by accomplishments in engineering (“without engineering there is no science”). Shadbolt also discussed the most recent manifestation of this marriage of science and engineering – the rise of AI – reviewing the roles AI is playing in science and how it is augmenting the scientific process.
Standing on the edge of error
Shadbolt emphasised that science has transformed the world across many centuries but we need to understand that it’s always a work in progress, “error is a persistent phenomenon and we need to manage and understand uncertainties”.
Taking us back to 1543, he pointed to two publications that changed the world – Andreas Vesalius’ seven-volume De Humani Corporis Fabrica Libri Septem, still considered one of the most influential books on human anatomy. It set out meticulous dissection as the basis for modern anatomy correcting errors based on animal anatomy. But, for Shadbolt, what made it revolutionary was the insistence on empirical evidence and hands-on research – the principles of the modern scientific method. “It also exemplified the spirit of the Renaissance in the bringing together of science and art, and would not have been possible without advances in art, engineering and printing.”
In the same year, the publication of Nicolaus Copernicus’s De revolutionibus orbium coelestium formulated a model of the earth revolving around the sun “which challenged classical views and the idea of humans being at the centre of the universe” and laid the groundwork for the work of Kepler, Galileo, Newton and many others.
“One century later Galileo was consolidating the role of systematic experimentation, mathematical analysis and an approach to understanding the world by carefully observed and repeatable experiments. In 1687 Newton’s Principia detailed his three laws of motion and his law of universal gravitation. It laid the foundations for classical mechanics, explaining Johannes Kepler’s laws of planetary motion, which Kepler had first obtained empirically – and making sense of the observations of Galileo, Copernicus, and others.”
Shadbolt emphasised that these contributions among others led to the ascendancy of the scientific method. “In a quest for truth and open mindedness, the scientific method builds on its own limitations, uncertainty and the possibility of error. It’s about questioning, testing, verifying, reviewing and repeating.”
“Science must always stand on the edge of error,” he continued. “It’s based on observation. The data may be wrong, there are instrumental limitations, human biases, sometimes errors are propagated, and ethics may not allow all types of experimentation. But there is built-in correction of our hypotheses via experimentation, peer review, and the expectation that a result should be replicable and reproducible. There is always acceptance that you may be wrong and may correct, and future generations may improve on your work.”
AI’s long history
Turning to AI, Shadbolt pointed out that it has a longer history than we may realise. The term was first coined at the Dartmouth Workshop in 1956 but had its origins in the work of British mathematician and computer scientist, Alan Turing, in particular his 1950 paper ’Computing Machinery and Intelligence’, which appeared in the philosophy journal Mind.
The original proposal for the Dartmouth Workshop held that any feat of human intellect could be so accurately described that a machine could do it.
Shadbolt described the combination of science, engineering, AI, human creativity and the massive escalation in computing capacities as having led to many applications of AI in science – such as proposing the 3D structure of proteins, exoplanet detection via giant telescopes, and analysing complex environmental data “that were not imagined or materialisable only decades ago”.
He described the stages of AI development from Symbolic AI in the 1980s which relied on the encoding of knowledge as rules and explicitly capturing human knowledge to solve particular tasks; Sub-symbolic AI in the 2000s which was about building neural networks and was biologically inspired and adaptive, and led to an extraordinary range of tools; to Generative AI in the 2020s which is capable of generating text, images and other data based on learning the patterns and structures of the input data, creating an AI ecosystem of tools and techniques.
“ChatGPT, for example, can ingest huge amounts of data and find patterns between words at very large distance. It is essentially a device for predicting the next most likely word in a sequence, which applied repeatedly becomes a coherent, extended response,” he added.
These tools will have a huge impact on science with, for example, their ability to rapidly conduct literature reviews (in one example reading 200 000 papers over lunch!), make relevant paper selections, and extract information.
“Such tools promise real advances for scientists,” he continued. “But they are not infallible and can be fooled. They are subject to endogenous error, use of outdated facts, information and reasoning failures, hallucination, confabulation. At their best performance about 4% of the facts in their summarisations are made up. They can be used to perpetuate scientific fraud as the information base can be flooded with machine-generated false results. We need to be able to prevent manipulation and make the technology more resilient.”
The explosion of such applications has also turned the field into an arms race with companies trying to out-do each other. “These companies say they don’t want to do harm, but the harm may not be anticipated. That’s why it’s important to have rules.”
There is no AI without data Shadbolt is on record as saying. “Data is a new kind of infrastructure and applications we haven’t even thought of yet will emerge. We must decide on the degree of uncertainty tolerated and what is absolutely required. Hence the need for ethics, regulation and governance. But decisions affecting a lot of people should involve a lot of people. Many minds need to be on it.”
“AI can generate spurious as well as accurate output,” he added. “It has the potential to be weaponised to undermine science, but equally can be used to confront disinformation. High-quality data plays a crucial role, and we can use AI to generate new insights.”
And Shadbolt hopes to put some of these ideas into practice in his STIAS project which looks at the use of open data and generative AI in the urban renewal Adam Tas Corridor project.
“This is an extraordinarily impressive development that is imagining new ways to connect space and place,” he said. “It needs the right kind of data stewardship with FAIR (Findable, Accessible, Interoperable, and Reusable data) principles at the heart. We need both to improve digital literacy and to ensure individual agency over personal data. It’s a wonderful vision, and a huge privilege to be part of it. I hope my background can be of use to move the vision forward. Generative AI offers a means to pull this data together, access and analyse it.”
And does he believe AI will ever fully replace humans?
“No, I am an optimist, with due care and attention AI can be used to empower and not oppress us. Will AI take all our jobs? No, humans are extraordinarily creative about inventing new things to do and be paid for (for example, social influencers didn’t exist a few years ago). There will be a huge need for human content creation in an age of AI.”
Michelle Galloway: Part-time media officer at STIAS
Photograph: Ignus Dreyer