Can AI produce scientific theory?

Essayist Axel Cypel, winner of the 2024 Roberval Prize, was asked during the ceremony whether AI could generate scientific theories. He contrasts the induction used by machine learning models, which generalize from existing data, with the hypothetical-deductive method of hard sciences. According to him, major physical theories like Darwin's evolution or Einstein's relativity were built by predicting phenomena before their observation, unlike AI's interpolation approach.

Cypel illustrates this with examples showing how scientific theories act as an "ontological winch," pulling new concepts into existence with minimal data. AI, however, produces models through high-dimensional interpolation, generating new outputs but not true conceptual creation. The underlying question touches on the nature of creation itself, whether scientific or artistic, and whether AI can ever cross that threshold.

The article also notes a note from the editorial team mentioning the refutation in October 2026 of the exponential time hypothesis by OpenAI and Anthropic, though this is unrelated to the debate on scientific theory.