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

MIT researchers tell WIRED that AI may soon become the default approach for solving theoretical problems in economics, high-energy physics, and other fields.

Last week, I attended a talk at MIT by Neal Thompson, who studies advances in computing and artificial intelligence. In just half an hour, Thompson made a compelling case that several scientific fields are about to be transformed by artificial intelligence.

While mathematicians are overwhelmed by the way artificial intelligence is automating their field, Thompson's latest study suggests that probability, statistical theory, economic theory, high-energy theory, gravity and cosmology, and quantum physics may be next, thanks to the rapidly evolving capabilities of frontier models. The study shows that between mid-2024 and mid-2026, the best AI models will go from being able to correctly solve about 2 percent of theoretical problems in these fields to about 35 percent of such problems. And there is no sign of this progress slowing down.

"Within a few years, we expect many theoretical questions to change completely in terms of what AI can do," says Thompson.

More generally, the Thompson Lab looks at how advances in computer science affect society and influence innovation. The study — which Thompson conducted with doctoral student Gabriel Manso and research scientist Claudio Meyrink Verdon — examined the problems presented in research papers posted online between December 2025 and May 2026, and then tested whether different OpenAI models—o1, o3, GPT-5, GPT-5, GPT-5, and GPT-5-5—solved them. they

Importantly, solutions to these problems were published in papers after training for GPT-5.5 and earlier models, which meant that the models they tested had to find their own solutions, rather than relying on answers provided as training data. (Such articles served as ground truth for testing the models' solving skills.) The researchers also blocked Internet access, which prevented the models they tested from searching for solutions online.

The areas covered by the study were narrow—the trio focused on areas where AI models could understand things without the need for simulation or testing. But this work suggests that the kinds of shockwaves currently being felt in mathematics are likely to spread to other math- and theory-heavy areas of scientific inquiry.

Last month, OpenAI shocked the math world when it announced that an unpublished model had solved a long-standing grand challenge in mathematics called Navier-Stokes, which deals with the equations that describe the behavior of fluid motion. The episode became more controversial after a mathematician at New York University — working with another academic associated with Anthropic — accused OpenAI of surprising them after learning of their progress.

OpenAI and others have since announced hundreds of other solutions to old math problems, sparking a crisis in the field. Some mathematicians have called for limits on the use of artificial intelligence, warning that the process of solving problems is often as useful as having the final answer.

As with mathematics, Thompson says we should expect progress in other areas, positive and negative effects. He notes that theoretical advances sometimes lead to practical advances in science, and suggests that more use of artificial intelligence may mean more progress in this way.

Some leading AI researchers have already established new labs focusing on AI for science, and major AI labs are all ramping up scientific research. These efforts explore how AI can be empowered to perform real-world experiments alongside theoretical analyses. Artificial intelligence may mean theorists no longer need to solve key problems, but researchers say that doesn't mean experts aren't needed to find solutions and make sure they're correct.

"We still need people to come up with good ideas to connect the dots," says Manso. For very complex problems, AI can make mistakes, he adds. (OpenAI actually retracted several solutions in its latest post-release dataset.) "We also need ways to verify what's been done, because it's very easy for a model to tell you, 'This is the right answer,' and when you look at the whole solution, you realize that maybe it's not."