Cite Forward

A Global AI Disclosure Standard Is Being Written Right Now. You Have Until 16 October to Shape It.

Every publisher has its own AI disclosure rules. They contradict each other, they are enforced inconsistently, and — on the available evidence — they do not change behaviour.

A serious attempt to fix that is underway, and it is currently open to public input. Most researchers have never heard of it.

What is being built

Five organisations are jointly developing a Global Reporting Standard for AI Disclosure in Research, intended to work across disciplines, regions, and publishing models:

It was proposed by Kari D. Weaver of the Ontario Council of University Libraries and Bert Seghers, president of the European Network of Research Integrity Offices, and formed the focus track of the World Conference on Research Integrity in Vancouver, 3–6 May 2026.

The output will be called the Vancouver Standard. COPE states the project aims to complete by the end of 2026.

It is not in force. Do not cite it as though it is.

Worth stating plainly, because a name like “the Vancouver Standard” invites the assumption that it already governs something.

It does not exist yet. It is in staged consultation. Nothing about it binds any journal today, and no publisher currently requires compliance with it.

The schedule, as published by the ISC:

  • Round 1 — December 2025 to February 2026 (closed). Mapping community needs and preferred disclosure structure.
  • Round 2 — July to October 2026 (open now). Thresholds, taxonomy, whether disclosure should be mandatory, and how to describe oversight and verification.
  • Round 3 — late 2026 to early 2027. Refining a draft standard based on feedback.

Round 2 opened on 9 July and closes on 16 October 2026. Contributions go through a webform at council.science/AIdisclosure, and organisations are explicitly invited to submit collective responses.

The questions actually on the table

Round 2 is the substantive one, and the issues it covers are the ones researchers argue about constantly.

Thresholds — when does AI use become disclosable? Grammar checking almost certainly does not require disclosure. An AI-generated first draft almost certainly does. Everything in between is currently a guess, and different journals draw the line in different places.

Taxonomy — what categories should exist? “AI was used” is close to meaningless. A controlled vocabulary distinguishing literature search from data analysis from drafting from translation would make disclosures comparable rather than decorative.

Should non-empty disclosure be mandatory? That is, must every paper state something — even “no AI was used” — rather than silence being ambiguous between no use and no disclosure.

How should oversight and verification be described? This is the most interesting one. It asks not just what tool you used, but what you did to check its output.

Why that last question matters more than the others

Disclosure alone has a weak track record. Analysis of more than 5 million papers found no measurable difference in AI-assisted writing between journals with disclosure policies and journals without them.

And a labelling regime has a structural limit. Telling a reader that a tool was involved says nothing about whether its output was checked. We covered this in the context of autonomous research agents, where the violations were invisible in the final manuscript — disclosure would have flagged the tool and revealed nothing about the problem.

A standard that asks what verification you performed is doing something different from one that asks what tool you used. The first creates a record that can be evaluated. The second creates a label.

The ISC’s own framing points that way — describing the goal as encouraging genuine reflexivity, mindful documentation, and meaningful disclosure. That is more ambitious than a checkbox, and harder to get right.

Why participating is worth twenty minutes

Standards get written by whoever shows up. If the responses come predominantly from publishers and integrity offices, the result will reflect what is convenient to administer. If working researchers contribute, it will reflect what is workable in practice.

The people most affected by a disclosure threshold are the ones who will have to apply it to their own manuscripts, in fields where AI use looks nothing like it does in the fields best represented on the drafting team.

There is also a regulatory reason to care. The EU’s Article 50 transparency obligations already carry legal force, and other jurisdictions are moving. A coherent international standard is the realistic alternative to a patchwork of incompatible national requirements — but only if it is good enough that regulators point at it.

If you want to contribute

  • Webform: council.science/AIdisclosure — expand “Consultation round 2”
  • Deadline: 16 October 2026
  • Collective responses welcome — departments, labs, and societies are encouraged to discuss internally and submit a summary

Round 3 will circulate a draft. By then the structural decisions — thresholds, taxonomy, whether disclosure is mandatory — will largely be made. Round 2 is where those get settled.

Whether the Vancouver Standard turns out to be useful or another ignored policy document depends substantially on who bothers to respond in the next eight weeks.

Sources

Dates and details are as published by the organising partners. The standard is in consultation and has not been adopted — treat it as a work in progress, not current policy. Research-based rather than hands-on — see our Methodology page.