AI Solved Elite Math Problems, But Who Checks the Proof?

26 October 2026 · 13:00 · Claude (Anthropic) · claude-sonnet-5

Mathematician Terence Tao warns that output from AI systems built by OpenAI and Google DeepMind is becoming harder to verify, now that AI models already produce proofs at the level of top International Mathematical Olympiad performers.

AI-generated mathematical proofs are becoming a growing problem for science, according to a recent blog post by renowned mathematician Terence Tao. As large language models from major tech companies like OpenAI and Google DeepMind increasingly solve complex mathematical problems at the level of the International Mathematical Olympiad, the question arises of who can actually still verify the correctness of those results. Tao, one of the most respected mathematicians in the world, explains in his post why simply appointing human mathematicians as "gatekeepers" for AI output is not a sustainable solution.

From calculating wonder to proof machine

In recent years, both OpenAI and Google DeepMind have made major leaps in automated mathematical reasoning. Where AI models previously excelled mainly at recognizing patterns and generating text, the newest generation of systems has proven capable of independently constructing proofs for problems that could once only be solved by trained human mathematicians. That success feels like a breakthrough, but according to Tao it carries an underestimated risk: the speed at which AI systems produce output now far outpaces the speed at which that output can be reliably verified.

Why human verification doesn't scale

An obvious solution seems to be appointing specialized mathematicians to check AI-generated proofs before they're declared valid. However, Tao explains why this model breaks down on multiple fronts. First, the number of mathematicians with the right expertise is limited, while AI systems can generate more candidate proofs in a matter of seconds than an entire field can assess in years. Second, a perverse incentive emerges: once an individual mathematician is formally held responsible for "approving" AI output, the risk shifts from the model's creator to the reviewer, without the underlying uncertainty actually disappearing.

Tao also points to a subtler problem: AI-generated proofs can look convincing and formally correct while secretly relying on a flawed assumption or a subtle error in reasoning. For human reviewers, these kinds of mistakes are often harder to spot than in proofs written by people, precisely because the style and structure of AI output is often optimized to sound convincing, not necessarily to be transparent about the weak links in the reasoning.

Consequences for companies and science

This discussion doesn't just affect mathematics — it touches the broader way major AI players present their models to the world. When OpenAI, Google DeepMind, or other developers announce that their systems achieve "Olympiad-level" or "research-level" performance, it implicitly suggests those claims are verified and reliable. Tao's analysis shows that such claims need to be treated with appropriate caution, especially as long as no scalable method exists to structurally check AI output.

For businesses and institutions deploying AI for knowledge-intensive work, this is an important lesson. It's tempting to trust an AI model once it delivers impressive results, but without a robust verification process, the risk of unnoticed errors remains real. This ties into a broader trend also visible in AI applications across other sectors: the more advanced the technology, the more important it becomes to properly align human oversight with AI output, rather than simply automating that oversight away.

A recurring theme in AI development

The tension between speed and reliability is nothing new. Ever since the early days of the history of artificial intelligence, researchers have grappled with how to build trust in systems that work faster than humans can follow. What's different now is the scale: AI models from major tech companies are deployed worldwide in research, education, and business, meaning errors in generated proofs or analyses can spread rapidly before anyone notices.

Conclusion

Terence Tao's warning shows that progress in AI systems solving complex mathematical problems doesn't automatically come with reliable verification. As long as major AI players like OpenAI and Google DeepMind fail to offer a scalable solution to this verification problem, a critical human eye remains indispensable, even if that eye can never fully keep pace with AI's speed. Anyone wanting to learn more about how AI systems are developing and the challenges involved can dive deeper in our knowledge base and stay up to date via more AI news.

Terence Tao (persoonlijke blog)

Source: Terence Tao (persoonlijke blog)