Yorkshire Philosophical Society

Event Information

Discoveries without Understanding

Dr Claire Malone

As artificial intelligence (AI) tools become embedded in scientific research – analysing data, proposing hypotheses, even writing papers – we are beginning to rely on systems that may lead us to answers faster and more efficiently than ever before.

But what happens when we don’t understand precisely how those answers were reached? Can science, which has long valued explainability and transparency, embrace a collaborator that often produces results without fully revealing how they were reached?

Despite their impressive results, the use of these AI tools raises some profound philosophical questions: Is AI merely extending the power of human curiosity – or is it reshaping it? Can an algorithm that finds patterns really discover something new? Or is AI at its most powerful when it works in tandem with human scientists, combining machine speed with human judgment, creativity, and doubt? If discoveries could come from systems we do not fully understand, what does it mean to ‘have trust in science’?

Join science communicator Claire Malone, a leading voice on the intersection of AI and ethics, as she examines the issues.

An event for York Festival of Ideas – free event booking essential with tickets from May 1st at 

https://yorkfestivalofideas.com/

7pm at The Auditorium, Bootham School, York, YO30 7BU

Member’s report:

What is Science ? This was a question posed by the speaker near to the start of the lecture. One starts with a hypothesis and then enters an iterative refinement process of observation and analysis. One then arrives at a conclusion and communicates it.

Claire introduced us to the ideas of the Vienna Circle who pioneered logical positivism. The circle argued that meaningful claims must be logically verifiable or empirically testable. She also discussed the ideas of Karl Popper who wrote the book The Logic of Scientific discovery in which he disagreed with the Vienna circle, promoted the idea of falsifiability and proposed constantly testing against reality. Science needs to be tested, challenged and there must be transparency. Karl Popper also asked  “Can a machine participate in the process ?”

Claire then introduced aspects of Artificial Intelligence (AI) which are systems that normally require human cognitive ability. Machine learning is a subset of AI and learned through two principal routes – supervised learning and unsupervised learning. Supervised learning depends on a human to initially characterise or label the data. With unsupervised learning, the machine attempts to identify clusters by grouping similar things together. A member of the audience was asked to put into groups a number of items. These were sorted according to material. In another test, the objects were more difficult to classify into neat groups. This provided a nice example of the challenges that AI faces as it is not always clear how best to classify the data that it is presented with.

A landmark research paper in 2017 called Attention is all you need introduced the transformer architecture. A sentence is split up into tokens which can be words, parts of words or punctuation. An encoding and decoding process can be used to determine the most likely word which follows in a sentence or the most likely next pixel in an image. AlphaFold is an AI application which can predict protein folding. AlphaFold shows that AI can solve major scientific problems. Demis Hassabis and John Jumper were awarded one half of the 2024 Nobel Prize in Chemistry for their work on AlphaFold.

At CERN AI is used to process large data sets and as a sophisticated filter, but the role of AI has begun to shift. Now the behaviour of the detector itself. Is modelled. A diffusion model reverses the process of adding noise. Generative Diffusion models are tens of times faster than traditional methods and are used for anomaly detection and searching for the unexpected.

Claire then asked “Is Science about understanding or producing patterns that work?” The AI Scientist can automate the scientific process, but AI searches within known boundaries. How close are we to machines doing science on their own? It seems that AI is starting to look like a collaborator. It is not an oracle that bypasses the scientific method; for scientists, it is a tool.

Chris Walker