AI Came for Math. Now It’s Chasing the Rest of Science

MIT professor Neil Thompson delivered a compelling lecture on the looming impact of artificial intelligence on various scientific disciplines. His recent study indicates that fields such as probability, statistics, economic theory, high-energy theory, gravitation, cosmology, and quantum physics are highly likely to be disrupted by AI advancements.

Over the span of just a few years, from mid-2024 to mid-2026, AI models have demonstrated remarkable progress, solving approximately 35% of theoretical problems in these fields, a significant leap from the initial 2% capability. The researchers tested leading AI models from OpenAI, including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5, on problems presented in papers published between December 2025 and May 2026.

These models had to independently find solutions, without relying on pre-existing answers trained into the models. Crucially, this test excluded internet access, ensuring the models had to generate solutions from scratch. The study focused on fields where AI could solve problems without requiring simulations or experiments, suggesting that AI's disruptive potential could extend beyond math to other theory-heavy scientific areas.

The announcement of OpenAI's AI solving the Navier-Stokes equations, a longstanding challenge in fluid dynamics, has stirred controversy and intensified a crisis in the mathematical community. Some mathematicians advocate for regulating AI's role in problem-solving, emphasizing the value of the problem-solving process alongside the final answer.

Despite potential negative consequences, Thompson emphasizes that AI's progress could lead to both positive and negative outcomes in science. He points out that theoretical advancements often drive practical innovations, and increased AI utilization may expedite such progress. AI labs dedicated to scientific applications are emerging, and major AI firms are intensifying their scientific research.

While AI may reduce the need for theorists to solve complex problems, experts will still be essential to interpret and validate AI-generated solutions. Manso stresses the continued necessity of human involvement in generating innovative ideas and ensuring the accuracy of AI-driven discoveries, as models can make errors.