Mathematicians warn AI could flood the proof economy with confident errors faster than humans can check them

Plausibility can become an attack on peer review

The declaration warns that models can produce arguments that look correct while containing subtle failures. Because drafts are cheap to generate, a flood of false claims can consume reviewer time and contaminate later work if errors pass into the literature.

This is not merely hallucination at a larger scale. Mathematics depends on a chain of independently checkable reasoning. When generation accelerates but qualified checking does not, the bottleneck moves from discovery to verification.

A commercial announcement is not community validation

The declaration criticizes mathematical achievements announced through press releases or blogs without the disclosure needed for scientific evaluation. Ars reports concern that proprietary models, undisclosed prompts, training data, and compute can make a dramatic result difficult to interpret or reproduce.

A correct answer may still be valuable. The scientific meaning depends on how it was produced, which human work it used, what resources were required, whether others can test it, and whether a narrow task is being sold as general reasoning.

Preserve the profession that can tell when the machine is right

The declaration also warns about attribution, copyright, unequal access to tools, distorted hiring incentives, and technology-company influence over research priorities. Early-career mathematicians may bear the largest cost if recognition moves toward proprietary systems and the questions those systems can solve.

The remedy is not nostalgia. It is stronger disclosure, responsible authorship, peer review, author rights, public compute infrastructure, and professional standards for AI-assisted results. The people capable of verifying a machine-generated proof are part of the safety system.

• Disclose tool use, prompts, methods, and human contributions at a reproducible level.
• Keep authors responsible for correctness and attribution.
• Support public compute and independent research access.
• Do not use commercial launch timing as a substitute for peer evaluation.