Terence Tao: AI 'Black Box' Proofs Are Polluting Mathematical Discovery
The Fields Medalist warns that corporate AI is solving open problems without providing the human understanding necessary to advance the field.
Fields Medalist Terence Tao has warned that AI companies are harming mathematics by producing rapid, opaque discoveries that bypass human understanding. Tao argues that when corporate labs provide answers without the underlying logic, they strip mathematical problems of their primary value as drivers for scientific advancement.
According to reports from New Scientist and Nowosci.ai, Tao believes the mathematical bottleneck has shifted. The challenge is no longer the generation of proofs, but rather their verification, exposition, and the ability of humans to comprehend them. He cautions that "black box" solutions from AI firms "pollute" open problems, effectively settling them without contributing new theories or methods to the community.
One concrete example of this trend is the progress made on the twin prime problem. OpenAI's GPT-6 Astra reportedly used Lean formalization to shrink the upper bound on the gap between consecutive primes to 186, down from the previous 246. While the result is a technical achievement, it exemplifies the risk of a result being correct yet incomprehensible to the researchers who would normally derive insight from the process.
The Erosion of Insight
Tao's perspective on AI has evolved over time. He initially viewed early models as "mediocre graduate students," but now considers AI a "junior research co-author." However, the integration of formal proof assistants like Lean has changed the stakes. The risk has moved from AI providing incorrect answers to AI providing correct answers that no human can explain.
For Tao, the answer is not the most valuable part of a mathematical problem; the value lies in the process of solving it. This cycle of hypothesis, testing, and rejection is what generates the new methods and theories that allow the field to grow. If this process is internalized within a corporate AI and only the final result is published, mathematics risks becoming a series of "checkboxes" where problems are solved but no new human insight is gained.
A New Framework for Prestige
To combat this "proof overload," Tao proposes three fundamental changes to how the mathematical community operates. First, he calls for the mandatory disclosure of AI use and the release of all associated logs. Second, he suggests shifting professional prestige away from being the "first to solve" a problem and toward being the "best to explain" the solution.
Finally, Tao advocates for requiring live presentations of results to experts to ensure the solver truly understands the logic. He warns that if the community does not ask these questions and set these standards, the future of the field will be settled for them by technology companies.