Artificial intelligence is getting increasingly good at solving mathematical problems. A few days ago, OpenAI reported that one of its models had solved the Navier-Stokes problem, one of the seven Millennium Prize Problems – a claim has been met with criticism by many mathematicians. With each new conjecture proven, it is natural to wonder whether AI's capabilities can extend to the sciences. Sociologist of science and Fellow of the British Academy Harry Collins disagrees – an essential part of doing science involves being part of a community that can invent new language to describe the unforeseen problems facing them. Without such an involvement with practising scientists, AI can never make a genuine scientific discovery.

The belief that artificial intelligence can make scientific discoveries is longstanding. For instance, it was claimed in the 1980s that an early AI program called BACON had rediscovered, among other things, Kepler’s Laws of planetary motion. But what BACON was doing was finding a pattern in data that already fitted Kepler’s Laws, not manipulating raw data to discover the patterns that the laws describe, let alone collecting the data itself. The latter requires engagement with the world and the separation of data from noise. The tension between finding patterns in agreed bodies of data and agreeing on what counts as data in the first place still besets all claims that computers can make scientific discoveries.

If the computer is working in a closed space, either because it has been fed clean data, or is working with models, or is playing games with fixed rules, it can make discoveries that humans have already made or have not made yet—even those that humans cannot make because they don’t possess the endurance, the speed or the calculating power of computers. That is how it comes to be that computers can now beat humans at games like chess and Go, and how computers can be helpful in analyzing medical X-rays, doing calculations of every kind, solving mathematical problems and even doing discovery work in certain kinds of science—a prime example being the protein folding for which the Nobel Prize was awarded recently.

SUGGESTED VIEWING AI and the mysteries of reality With Roman Yampolskiy, Steve Fuller, Suzanne Livingston, Güneş Taylor

Molecular biology is what I call “hypernormal” science because the puzzles it solves are a matter of exploring more and more of a domain comprising fixed entities and relationships over which scientists no longer disagree but simply do not have the resources to match the machines—as in chess and Go. The trouble is, not all science is like this; a lot of it is deciding on what to look for and what to manipulate and what counts as a result, just like what Kepler was doing when the laws of planetary motion were first discovered or, as we should say, “established”. To do that kind of science, one needs to mimic human intelligence, not just calculate better and faster. Unfortunately, human intelligence is not like computer intelligence. The latest generation of computers are doing an almost unbelievably good job at mimicking aspects of human intelligence, but the difference is still there at the frontiers for those who know how to look.