Responsible Release of AI-Generated Mathematics

In the rapidly evolving landscape of artificial intelligence, a growing concern has emerged: the release of AI-generated mathematical results. Some advanced mathematical problems are being tested on proprietary models that remain inaccessible to the wider scientific community. Responsible release of such AI-generated mathematics is crucial, as human understanding and oversight remain paramount.

The mathematical community has long held the belief that the dissemination and peer review of results are essential for the reliability, trustworthiness, and effectiveness of mathematical work. A fundamental norm in this field is that authors must fully understand the mathematical argument, verify its correctness themselves, and assume full responsibility for the content. Furthermore, mathematicians often share their work through seminars and conferences, fostering a deeper collective understanding.

However, with the advent of AI, it is now possible for machines to generate mathematical arguments without the human prompting them being able to comprehend, verify, or take responsibility for those arguments. This raises significant questions about how to navigate the new paradigm of AI's role in mathematical output.

To address this issue, a group of mathematicians and AI experts conducted a survey, garnering over 600 responses from the mathematical community. Based on these insights, they have formulated a set of recommendations for AI labs to responsibly release mathematical results. The overarching principles behind these recommendations emphasize the importance of human understanding and responsible academic practices.

If the AI-generated mathematical results are comprehensible to a responsible mathematician, they should adhere to the traditional norms of the mathematical community. This includes posting preprints, submitting papers for peer review, and giving talks to explain the work to colleagues. For cases where the AI output is not fully understood by the prompting mathematicians, two courses of action are recommended.

Firstly, the labs should follow technical norms for the release of AI-generated mathematics, ensuring proper citation of related ideas and producing well-written, comprehensible proofs. Secondly, AI labs should provide support for additional mathematical activities necessary for humans to understand and assimilate the AI-generated content, including the identification of possible applications.

Key recommendations include timely depositing results in open, scholarly repositories with persistent identifiers and comment functionality. AI labs are urged to refrain from using mathematical result releases as mere marketing tools for their models. They should also provide public information about the models used, prompts, chain of thought, time taken, and estimated computation costs. Ideally, the released proofs should be formalized according to community standards.

By adhering to these guidelines, AI labs can contribute to the development of a new paradigm that incorporates human understanding and scholarly rigor into the release of AI-generated mathematical results.