150 mathematicians told governments not to believe AI hype. The industry responded by redefining the hype.

The Pattern

In 2026, over 150 mathematicians from around the world signed the Leiden Declaration on AI and Mathematics, warning governments not to “believe the hype” when it comes to AI’s capabilities to solve complex mathematical problems. The declaration was 11 pages. The signatories included the vice president of the International Mathematical Union and the head of computer science at the University of Oxford.

The core warning was simple: “There is currently a strong commercial incentive on the part of the technology industry to overstate the capabilities of their products.” The declaration advised policymakers to “consult with experts, including mathematicians, in forming policy decisions rather than relying on press releases or popular reporting of mathematical results.”

The plausible-but-wrong problem

Leslie Ann Goldberg, a signee and head of computer science at Oxford, identified the specific danger: “Current automated techniques can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs. This is a serious problem: research in mathematics (and in mathematical disciplines like theoretical Computer Science) depends on the ability to trust published results.”

This is the deeper problem with AI-generated mathematical output. It is not that the AI produces obviously wrong answers. It produces answers that look right, with the surface structure of a proof, the notation, the logical flow, but that contain subtle errors invisible to anyone who does not have the expertise to check every step. The output looks like mathematics. The process is entirely different. That difference is invisible to anyone reading only the final answer.

The Erdős problems

The declaration was prompted in part by two high-profile incidents. First, a 23-year-old without formal mathematics training claimed to have used ChatGPT to solve one of the Erdős problems, a database of challenging conjectures left by Hungarian mathematician Paul Erdős. Then, OpenAI claimed its AI had disproved an 80-year-old “unit distance” conjecture, also devised by Erdős. Both claims generated significant press coverage. Neither survived scrutiny by working mathematicians.

The pattern is the same one documented across AI hype: a dramatic claim, heavy press coverage, no follow-up when the claim turns out to be overstated or wrong. The mathematicians who signed the Leiden Declaration are not anti-AI. They are anti-hype. They are saying what every working scientist knows: the difference between a result that looks correct and a result that is correct is the entire difference. AI is very good at the first. It has not demonstrated the second.

Why this matters beyond mathematics

The Leiden Declaration is about mathematics, but the problem it identifies is general. If AI can produce plausible-but-wrong mathematical proofs that are difficult to distinguish from correct ones, the same is true for legal arguments, medical diagnoses, policy briefs, and every other domain where the output looks like expertise. The danger is not that AI will replace experts. The danger is that AI will produce outputs that look like expert work to people who cannot tell the difference. The trust follows the appearance. The correctness does not.

This is the propaganda function of AI hype: it takes the surface appearance of competence and presents it as competence itself. The Leiden Declaration is 150 mathematicians saying: do not confuse the two. The industry’s response was to redefine AGI, extend deadlines, and generate more plausible outputs. That is the proof they heard the warning. They just have no incentive to act on it.

Verdict: Misleading. The AI industry’s claims about mathematical capability are overstated. 150 mathematicians said so. The industry’s response was to keep making the claims. The hype is the product. The mathematics is the evidence that the product does not work as advertised.