AI Came for Math. Now It’s Chasing the Rest of Science
Researchers say AI is beginning to solve problems in several theory-heavy sciences, suggesting that changes already affecting mathematics may spread to other fields. A study led by Neil Thompson found that leading AI models improved sharply in their ability to solve theoretical problems, though the researchers stress that human experts will still be needed to interpret and verify results.
The study tested OpenAI models on problems from recent research papers in probability, statistics, economics, high-energy theory, gravitation and cosmology, and quantum physics. Between mid-2024 and mid-2026, the models’ success rate rose from about 2 per cent to about 35 per cent; internet access was blocked, and the papers’ solutions post-dated the models’ training cut-offs. The researchers focused on problems that did not require experiments or simulations, and note that AI can make mistakes, with some published solutions later withdrawn.
- AI models’ success on selected theory problems rose from 2% to 35%.
- The study covered six areas beyond mathematics.
- Experts remain essential for checking and interpreting AI solutions.
New here? Start with this
AI systems can identify patterns and work through problems using methods learned from large collections of text and other data. In mathematics, researchers have been exploring whether these tools can help with proofs and difficult questions; several other sciences also rely on theory and mathematical reasoning.
The fields in this study include probability and statistics, which describe uncertainty and data, and economics, which examines choices and markets. High-energy physics, gravitation, cosmology and quantum physics use mathematical theories to explain matter, gravity and the universe.
AI-generated answers still need checking: a plausible-looking result may be wrong, and scientific claims often need careful interpretation. Researchers therefore see these tools as potential aids to experts, whose judgement remains important in deciding whether a result is sound and what it means.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
AI’s growing ability to solve hard theoretical problems could help researchers explore more ideas, check complex reasoning and make progress across fields where experiments are not always practical. If experts verify its work and remain responsible for interpreting results, it may become a useful research tool that expands human capacity rather than replacing it.
The case against
A rising success rate on selected problems does not establish that AI can reliably produce sound or significant scientific results across whole fields. Mistakes may be hard to detect, and withdrawn published solutions underline the need for caution; relying too heavily on AI could divert attention from verification and weaken confidence in research.