AI is trained on evidence and judgements produced within our existing systems of knowledge. Its estimate of what is plausible therefore reflects, in part, what we already recognise as plausible.

But AI can also detect patterns across quantities and combinations of data that humans cannot perceive. The sheer number of such connections is vast. Much of it may be noise, but it may also contain knowledge we do not yet have.

AI systems already generate hypotheses, test them and revise them in response to the results. OpenAI and Anthropic have recently reported AI systems producing new mathematical results, including advances on problems researchers have worked on for decades.

An idea may seem implausible because it does not fit our current concepts, assumptions or theories. However, apparent implausibility does not, by itself, distinguish error from novelty.

Suppose, for example, a system detects a weak but persistent relationship across thousands of variables that researchers had never thought to connect. The system can then test that relationship on independent data, explore alternative explanations and use the results to develop or reject the hypothesis.

An objection is that knowledge requires a knower and that AI does not qualify as one. On this view, AI can produce candidate ideas, but these become knowledge only when people have good grounds to believe them.

Even on that view, AI's contribution may be significant. It can inherit our existing knowledge and assumptions and, it seems, generate possibilities that extend beyond them. Some of which we may come to know.

OpenAI, Ten advances in mathematics and theoretical computer science https://openai.com/index/ten-advances-in-mathematics/

Anthropic, Learning more about Claude's mathematical capabilities https://www.anthropic.com/research/riemann-zeta