Mathematicians Are Saying Don’t Believe the AI Hype

Days after OpenAI announced that an internal model had disproved the Erdős unit distance conjecture, over 150 mathematicians from across the world signed a declaration telling governments not to believe the hype about AI mathematical capabilities. The Leiden Declaration, backed by the International Mathematical Union, warned that tech companies have a strong commercial incentive to overstate the capabilities of their products and urged professional mathematicians to resist beating the drum for AI developers.

The people closest to the field are pushing back while the AI companies and the media amplify the capability claims. That gap between insider skepticism and public narrative is worth examining carefully.

What the Mathematicians Are Actually Worried About

The Leiden Declaration covers several distinct concerns worth separating because they are not all the same argument.

The commercial incentive concern is the most fundamental. Tech companies have strong financial incentives to overstate AI mathematical capabilities because impressive mathematics results drive investment, talent recruitment, and regulatory deference. Kevin Buzzard at Imperial College put it pointedly: mathematicians should find it quite striking that tech companies are suddenly interested in their work. The interest is not disinterested.

The proof quality concern is the second major issue. AI systems can generate plausible-looking proofs that contain subtle errors that human reviewers miss, particularly under time pressure or when the reviewer is not a deep specialist in the specific subfield. The declaration specifically warns about flooding the field with plausible but flawed proofs that could corrupt the mathematical literature if not caught by rigorous peer review. Independent verification by specialists not affiliated with the company making the capability claim is the standard that needs to be applied consistently, and the declaration suggests it is not being applied consistently.

The attribution and credit concern is distinct from the capability question but matters for the sociology of the field. If AI systems are contributing to mathematical results, questions about authorship, credit, and the career incentives that drive mathematical research become genuinely complicated. A graduate student who proves a theorem gets a career. An AI system that proves a theorem gets a press release. The declaration is partly a professional community protecting its members from having their contributions devalued, which is a legitimate concern even if it is also somewhat self-interested.

Martin Hairer’s Assessment

The most analytically significant comment in the coverage comes from Martin Hairer, a Fields Medal winner and one of the most celebrated mathematicians working today. The idea of mathematicians being replaced by AI is complete nonsense in his opinion, he said, adding that he finds it hard to believe that the type of models currently available will suddenly start producing genuinely new insights.

Hairer is not saying AI cannot do useful things in mathematics. He is saying the mechanism is not genuine insight and the implication that mathematicians face structural displacement from current AI systems is not supported by what the systems are actually doing. The systems are doing more, faster, within the space of known mathematical techniques and connections. They are not generating genuinely new insights in the sense that transforms a field.

The Verification Problem Is Worse Than It Appears

One detail buried in the coverage of the Erdős result that the Leiden Declaration brings into sharper focus: the breakthrough reportedly came following a very simple chatbot inquiry. That framing, if accurate, raises serious questions about the human scaffolding involved in the result.

The difference between a model autonomously discovering a novel mathematical connection and a model responding to a carefully constructed prompt that guided it toward a known solution space is analytically enormous. The former is evidence of genuine autonomous capability. The latter is evidence of an effective human-AI collaboration where the human provided the crucial directional insight and the model executed within that direction.

The way AI mathematical results get reported, autonomously solved a prominent open problem, obscures the degree of human involvement in structuring the problem, directing the search, and verifying the output. The 150 mathematicians who signed the declaration are raising exactly this concern. Independent scrutiny of the human scaffolding question is essential before the capability claims can be accepted at face value.

The Silent Majority Problem

The Leiden Declaration is also a data point about the sociology of the AI discourse that extends well beyond mathematics. Yann LeCun recently observed that most leading AI figures think the existential-risk claims are complete nonsense, yet stay silent while the alarmists attract disproportionate attention. The Leiden Declaration is 150 mathematicians deciding not to stay silent.

The declaration explicitly notes that their intervention follows claims of increasing capability from AI firms. They are responding to the hype cycle directly. The International Mathematical Union is lending institutional weight to a pushback that individual mathematicians have been making in private but not in organized public form.

Buzzard’s observation that mathematicians should find it quite striking that tech companies are suddenly interested in their work is sharp and worth sitting with. Pure mathematics has historically been so abstract and removed from commercial application that industry interest was essentially nonexistent. Suddenly the biggest companies in the world are sponsoring mathematics competitions, funding mathematics research, and issuing press releases about mathematics breakthroughs. Understanding why requires understanding what those breakthroughs are worth commercially in the current AI narrative environment, which is a great deal regardless of their actual mathematical significance.

The Honest Picture

AI systems are genuinely producing useful mathematical outputs in domains where verification is automated and binary. The Erdős result, whatever the degree of human scaffolding involved, represents a real capability that did not exist five years ago. That is worth acknowledging.

But the systems are not producing genuine novel mathematical insights that transform the field in the way that human mathematical genius does. The Fields Medal winners signing the Leiden Declaration are not at risk of being replaced by current AI systems on any near-term timeline. Their work operates at a level of genuine novelty and creative synthesis that current AI architectures do not approach.

The honest picture is that AI is becoming a powerful tool for mathematical exploration in specific constrained domains while the highest gradient mathematical work remains deeply human. The Leiden Declaration is 150 mathematicians signing their names to that conclusion.

The silent majority spoke this week. They said the same thing they have been saying in private. The tool is useful. The hype is not.