Reference frameworks for the norm-building segment

Chair’s notes for the session corpus, 6 October 2026. The session proposal names the Leiden Declaration as “the obvious model” for a discipline-level response, and the norm-building segment asks for candidate norms in a fixed form. This note summarizes both. The declaration is summarized, not reproduced; read the original at the link.

The session’s norm form

Each candidate norm names three things:

  1. a concrete activity;
  2. the conditions under which AI involvement in it is acceptable;
  3. who is answerable for the resulting decision.

The panel accepts, revises or rejects each norm on screen. The proposal says the norms draw mainly on guiding questions 3 (what must be verified, documented and disclosed) and 5 (where authority and accountability remain human). The panel decided (6 October 2026) that the norms leave the room as a session artifact, not as a declaration.

The Leiden Declaration on Artificial Intelligence and Mathematics (2 June 2026)

  • A community initiative endorsed by the International Mathematical Union (DOI 10.5281/zenodo.20302944).
  • It grounds its recommendations in five values of mathematical research, summarized here:
    1. proofs confer certainty and understanding;
    2. results are attributable to authors who take credit and responsibility;
    3. arguments are transparent and independently verifiable;
    4. the community keeps shared standards for judging depth and significance;
    5. understanding and expertise develop within autonomous communities of mathematicians.
  • Recommendations to individual mathematicians include:
    • disclose the use of automated tools, including large language models;
    • make verification easier for peer review;
    • keep human responsibility for correctness;
    • affirm that authorship remains human;
    • attribute carefully;
    • choose tools in line with these values.
  • It also addresses mathematical organizations and funders (lead on publishing and reviewing policy, keep standards of rigor for automated results), policymakers, and commercial AI.

Points of comparison the panel may draw

These are the chair’s notes, not claims made by the declaration.

  • Verification. Mathematics can verify a proof. Measurement has no single equivalent; the closest is a validity argument, which is cumulative and judgment-laden. What would independent verification of an AI-assisted analysis mean in measurement?
  • Authorship and responsibility. The declaration keeps both human. Briggs’s question (whose judgment is a review revised after an AI critique?) and Betebenner’s (who answers for work beyond hand-checking?) test the same principle in measurement practice.
  • Disclosure. The declaration asks authors to disclose tool use. Briggs’s scan found 13 of 14 measurement journals require authors to disclose generative AI use. Reviewer-side rules are inherited from publishers.

Source: https://leidendeclaration.ai/