EGM AI sets a charter for AI-generated mathematics: labs must disclose costs and failure rates

The “EGM AI” initiative has set a detailed guidance framework that defines responsible publishing rules for mathematical results generated by AI models, after a survey that included more than six hundred specialists in the academic mathematics community, coming in response to OpenAI’s announcement of achieving numerous mathematical results without disclosing details. The initiative also announced its unequivocal rejection of testing advanced mathematical problems on closed models that are not available to the scientific community, urging leading labs to cease this practice immediately, as it threatens to isolate mathematicians from the trajectory of their field’s development.

The charter is based on a firm principle that the essential value of mathematics lies in human understanding of proofs and the researcher’s ability to explain, scrutinize, and assume scientific responsibility for them, which contrasts with the current reality of models that produce lengthy inferences and non-standard terminology that even the persons who formulate their commands cannot comprehend or verify.Publishing mathematical proofs is not a marketing platform for promoting models, but a scientific commitment that requires full documentation of the reasoning process and computing costs.The recommendations require labs to audit prior scientific output to attribute ideas to their originators even when the model arrives at them independently, and to present results in clear, understandable academic language instead of confusing formulations.

Phase one of the guidelines mandates depositing results in independent scientific repositories that are not under lab control, granting the research permanent citation identifiers and opening a venue for comments and review. The document requires labs to disclose the model name, the command text, a summary of the reasoning chain, the time taken, the estimated computing cost, and to publish the initial outputs before any refinement. It also stipulates providing visual specification files for the proofs according to accepted standards, such as including verification and comparison files and programmatic linking data, together with a systematic documentation of how the problem was selected and the proportion of similar problems the model attempted to solve and failed on.

In phase two, the recommendations oblige AI labs to fund human assimilation activities for those results through established independent non-profit organizations, including financing conferences, summer schools, and research teams, as well as supporting post-doctoral researchers, students, and the authoring of books and extensive commentaries, without this funding constituting a recognition of the labs’ practices.A scientific equation loses its legitimacy when open collective verification turns into a monopolised technical privilege.The document warns of the emergence of a two-tier research system in which the gap widens due to technical privilege, institutional wealth, and the monopoly of advanced models.

This shift reshapes work priorities for researchers and mathematics and computer science departments in universities and research centers across the Gulf, Egypt, and the Arab Levant. Imposing standards for visual proof and open code documentation gives regional research teams, which typically lack budgets for training massive models, the ability to compete in verification, conceptual analysis, and building on published results, rather than remaining passive observers of opaque commercial announcements. Requiring transparency in computing costs and reasoning chains also enables academics in the region to assess the actual viability of the tools offered and steer their research toward independent mathematical verification, as well as to participate in fellowship programs and international workshops whose funding, as stipulated by the document, must come from sources outside the direct dominance of the labs.