AI has made breakthroughs in mathematics, but far less progress in physics. “Physical intuition” is hard to formalise in training, and thus struggles to link maths to reality. Still, its strengths in derivation and data analysis let physicists focus on creativity and critical thinking.
By Yusen Wang
“I slumped in my chair, dizzy, as if witnessing an atomic bomb explosion.” This was how a Chinese mathematics PhD student reacted to the news that AI had apparently solved the Navier-Stokes equations.
On September 5, 2026, an internal OpenAI model spent 88 hours producing a proof that the three-dimensional Navier-Stokes equations, under smooth external forcing, can develop a singularity in finite time. GPT-6 Astra then spent a further 17 hours formalising and verifying the proof in Lean. On September 11, the Clay Mathematics Institute announced that the Navier-Stokes problem had “apparently been settled”. If confirmed formally, this would mean that one of the seven Millennium Prize Problems had been solved.
AI had already made progress on longstanding mathematical problems, including the unit distance conjecture and several problems posed by Paul Erdős. Although some results still require verification, there is no doubt that AI is advancing on problems once thought to demand the abilities of top human mathematicians.
Many therefore expect physics to be the next field transformed by AI. Yet in 2025, CritPt, a benchmark created by over 50 physicists using problems from their unpublished research, found that the best-performing model, GPT-5 (high), achieved just 4.0% accuracy, rising to around 10% with coding tools. A year later, the latest models reached about 32%, but still failed on roughly 70% of tasks. Although AI can indeed conduct certain physics tasks, its research capabilities remain far below those of human physicists. What, then, can it still not replace?
AI’s Tricks
Physics research involves formulating hypotheses, mathematical modelling, experimental verification, and interpretation. Much mathematical work, including derivations, equation solving, and function fitting, suits current AI systems. Having learned patterns of human reasoning, AI can explore multiple paths and prioritise promising ones. It resembles a slime mould—rather than relying on a single intuition, it explores many routes and retains those that work.
Mengdi Wang of Princeton University recently explains that mathematics and code have explicit verifiers. Lean checks proofs, while compilers check programs. These provide clear feedback on whether a proposed solution works, allowing AI to refine its reasoning and explore alternatives.
Unreasonable Intuition
The challenge changes when mathematics must connect to the physical world. Human physicists do not simply follow fixed procedures; they also rely on so-called “physical intuition”, developed through sustained engagement with mathematics and experiments. It helps them identify relevant features, decide what to measure, and connect results to physical meaning.
Zahavy argues that scientific discovery often requires a leap from observation to a new explanatory hypothesis. Physical intuition may therefore involve abductive reasoning that is difficult to formalise. Bisht et al. note that much tacit knowledge, including knowledge of failed approaches, is rarely documented and thus rarely appears in training data. If such knowledge helps scientists judge which ideas are worth pursuing, AI trained mainly on recorded science may struggle to identify promising explanations.
There is experimental evidence to support that AI lacks basic physical intuition. In a 2025 study testing whether multimodal models could judge the stability of block towers, GPT-4V and Claude-3 performed only slightly better than chance. A 2026 follow-up found that interaction improved performance on the same task, but the ability did not reliably transfer to related tasks.
Scaffolding or Explorer?
So what role can AI play in physics research? A 2024 perspective in Nature Reviews Physics concluded that theorists are unlikely to be replaced by AI in the near future, while human expertise combined with AI will become increasingly important. AI may serve as scaffolding, handling tedious mathematical work while humans identify important patterns and propose new ideas.
Wang’s team is developing automated experimental platforms that could make AI a more capable experimentalist. A 2026 paper proposes a “Reverse Interactive Theorem Prover” to formalise abductive reasoning: generate a hypothesis, derive its consequences, compare them with data, and revise it. Such systems could move AI beyond verifying solutions towards exploring scientific explanations.
Perhaps the more important question is not whether AI can replace physicists, but what will remain once tedious work is automated. The answer may be the ability to notice subtle details in the real world. Gravity, quantum mechanics, relativity—those great theories may have begun with something as simple as a falling apple, a curve that refused to fit, or a beam of light that could never quite be caught. Trust your intuition; it may be a kind of magic you can be proud of.
References
Alexeev, B., Barreto, K., Li, Y., Lichtman, J.D., Liam, P., Iqbal, S.J., Tang, Q. and Tao, T. (2026). Primitive sets and von mangoldt chains: Erd\h{O}s problem #1196 and beyond. arXiv. Available at: https://arxiv.org/abs/2605.00301v1 [Accessed 4 Oct. 2026].
Artificial Analysis (2025). CritPt: Score. [online] Artificial Analysis. Available at: https://artificialanalysis.ai/evaluations/critpt [Accessed 4 Oct. 2026].
Bishop, K. (2026). OpenAI claims to have solved the 90-year-old navier-stokes math problem in 88 hours. [online] CNBC. Available at: https://www.cnbc.com/2026/09/09/openai-navier-stokes-math-problem-solved.html [Accessed 4 Oct. 2026].
Bisht, H., Kumar, V., Jablonka, K.M., Mausam and Krishnan, N.M.A. (2026). Agentic AI scientists are not built for autonomous scientific discovery. arXiv. Available at: https://arxiv.org/abs/2605.08956v1 [Accessed 4 Oct. 2026].
Clay Mathematics Institute. (2026). Navier-stokes announcement - clay mathematics institute. [online] Available at: https://www.claymath.org/news/navier-stokes-announcement/ [Accessed 4 Oct. 2026].
Fefferman, C. (2026). Existence and smoothness of the navier-stokes equation the Euler and navier-stokes equations describe the motion of a fluid in R N. [online] Available at: https://www.claymath.org/wp-content/uploads/2022/06/navierstokes.pdf [Accessed 4 Oct. 2026].
He, Y.-H. (2024). AI-driven research in pure mathematics and theoretical physics. Nature Reviews Physics, [online] 6(9), pp.546–553. doi:10.1038/s42254-024-00740-1.
Noga, A., Bloom, T.F., Gowers, W.T., Litt, D., Sawin, W., Shankar, A., Tsimerman, J., Wang, V. and Wood, M.M. (2026). Remarks on the disproof of the unit distance conjecture. arXiv. Available at: https://arxiv.org/abs/2605.20695v1 [Accessed 4 Oct. 2026].
OpenAI (2026a). On the navier–stokes millennium prize problem. [online] OpenAI. Available at: https://openai.com/index/navier-stokes-solution/ [Accessed 4 Oct. 2026].
OpenAI. (2026b). An OpenAI model has disproved a central conjecture in discrete geometry. [online] Available at: https://openai.com/index/model-disproves-discrete-geometry-conjecture/ [Accessed 4 Oct. 2026].
Schulze, B.L.M., Voudouris, K., Can, D. and Schulz, E. (2026). Can vision language models learn intuitive physics from interaction? arXiv. Available at: https://arxiv.org/abs/2602.06033v2 [Accessed 4 Oct. 2026].
Schulze Buschoff, L.M., Akata, E., Bethge, M. and Schulz, E. (2025). Visual cognition in multimodal large language models. Nature Machine Intelligence, [online] 7(1), pp.96–106. doi:10.1038/s42256-024-00963-y.
Shalyt, M., Regev, N., Soljačić, M. and Kaminer, I. (2026). Can AI follow in Einstein’s footsteps? arXiv. Available at: https://arxiv.org/abs/2607.27794v1 [Accessed 4 Oct. 2026].
Zahavy, T. (2026). Position: LLMs can’t jump. [online] Icml.cc. Available at: https://ICML.cc/virtual/2026/poster/67091 [Accessed 4 Oct. 2026].
Zhu, M. et al. (2025). ‘Probing the critical point (CritPt) of AI reasoning: A frontier physics research benchmark’. Available at: https://arxiv.org/abs/2509.26574v4` [Accessed: 4 October 2026].
Li (2026). ‘Wang Mengdi: Admitted to Tsinghua University at 14 and Princeton’s youngest tenured professor — AI has yet to discover new fundamental science | Full speech included’. *InfoQ*. Available at: https://www.infoq.cn/article/d0NzJlH20aPXVmsZHdjW [Accessed: 4 October 2026].
Analysis Enthusiast (2026). ‘OpenAI announces breakthrough in solving the Navier–Stokes equations, one of the seven Millennium Prize Problems’. Bilibili. Available at: https://www.bilibili.com/video/BV1SdYn6cEoq/ [Accessed: 10 October 2026].
