A study presented at MIT by Neil Thompson and colleagues suggests AI is poised to disrupt several math-heavy scientific fields. The researchers tested OpenAI models on theoretical problems in probability, statistics, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. Between mid-2024 and mid-2026, the best models' success rate rose from about 2% to 35%, according to WIRED.
The test was designed to measure genuine problem-solving rather than memorization. The problems came from research papers posted between December 2025 and May 2026, with solutions published after the training cutoff for the models tested. The researchers also blocked internet access, so the models could not look up answers. The fields were limited to theory problems that did not require simulations or experiments.
The authors caution that AI progress will have both positive and negative effects. Theoretical advances can drive practical science, but AI solutions can be wrong; OpenAI withdrew several solutions from its latest dataset after publication. Manso says people will still be needed to connect the dots and verify what AI produces.