On September 29, 2026, the AGMI organization published a report titled 'general-sep29' that examines how frontier AI laboratories are testing advanced mathematical problems on proprietary models inaccessible to the wider scientific community. The report notes that such internal testing bypasses longstanding mathematical norms—such as authors verifying their own work and explaining results in seminars—raising concerns about reliability and trustworthiness when AI produces arguments that no human can understand or take responsibility for.
In response to over 600 replies from the mathematical community, AGMI proposes a two‑step framework for responsible release. First, during an initial release, labs should attribute related existing literature, prompt models to write proofs in conventional mathematical style with clear introductions and precise statements, deposit results promptly in open scholarly repositories that allow comments, disclose the model name, prompts, a summarized chain of thought, computation time and estimated cost, and, where possible, formalize the proof with community‑standard artifacts and machine‑readable metadata linking natural language and formal versions. Second, labs should support the development of human understanding by providing funding—through existing nonprofit institutions—for community‑led activities such as conferences, summer schools, targeted workshops, working groups, postdoctoral fellowships, and expository writing projects that help assimilate and build upon the AI‑generated output.
The report further warns that relying on proprietary internal models risks creating a two‑tier system in which a few labs outpace the rest of the field, alienating the mathematical community from its own discipline. To counter this, AGMI advises labs to grant the global mathematical community broad, equitable access to their publicly available models, stressing that progress in mathematics depends on collective verification, conceptual synthesis, and shared intuition, which cannot thrive when frontier instruments are limited to a select few. The document emphasizes that these recommendations are community‑informed proposals and does not indicate that any AI lab has currently adopted them or assesses their effectiveness.