Summary

  • The IGF says its Secretariat may use AI-assisted tools to organise or analyse IG Lab submissions, but not to evaluate proposals. “May” does not establish that a tool has been used, and “analysis” does not identify which outputs a reviewer will see.
  • The full call separates eligibility screening, review by a small multistakeholder group, a shortlist of up to six, portfolio balancing and MAG selection of up to three. A 3 September FAQ said the exact reviewer strategy was still being finalised after the submission pool became known.
  • I propose a proposal-level boundary receipt linking the original submission to every AI-assisted administrative action and every responsible human decision, including eligibility, merger, conflict recusal, shortlist, balance adjustment and final selection.

At 23:59 UTC on 15 September, the first Internet Governance Lab stops being only an invitation and becomes a queue. Each proposal will acquire labels, summaries, comparisons and a place in a review sequence. The institutional question is not whether software may help handle the queue. It is whether anyone can later show when handling ended and judgement began.

The call page draws the line in unusually direct language. The Secretariat may use AI-assisted tools for tasks such as organising or analysing submissions, but not for evaluating proposals. Applicants receive a related but separate rule: they may use AI for drafting or language refinement, while the proposal should retain their own ideas, expertise, policy questions and genuine human perspective.

Both statements are narrower than a general AI policy. The first does not say that AI has been used, which system might be chosen, what data it would receive, or which outputs would reach reviewers. It creates permission and a prohibition. Accountability begins in the space between them.

An administrative label cannot settle the boundary

Alphabetising files is plainly administrative. Applying a topic label may also be administrative, until the label determines which expert reads the proposal or which thematic stream has room. A similarity cluster may help staff find overlap, but it may also frame two independently developed proposals as candidates for merger. A summary may save time while changing salience: an omitted limitation can make a weak proposal look mature, and an omitted implementation detail can make a strong proposal look abstract.

None of this proves that IGF will use those functions or that an AI output will influence a decision. It shows why “organising or analysing” is not a self-executing safeguard. The test is downstream effect. If a machine-produced field changes the comparison set, reviewer assignment, order of attention or evidence visible at a decision, the human act that accepts that field must be recorded.

This is consistent with the voluntary NIST AI Risk Management Framework, which calls for defined human-AI roles, scoped use, documented oversight and traceable risk information. NIST does not govern the IGF process. It offers a useful comparison: saying that a human remains “in the loop” is weaker than naming the person, task, input, output and authority.

The funnel contains more than one judgement

The full call gives the process several distinct gates. The Secretariat first checks timeliness, completeness and word limit, thematic scope, whether the submission contains a sufficiently concrete governance proposal, and whether it is suitable for publication and multistakeholder examination. The last three checks are not mechanical facts in the way a timestamp is. They require interpretation.

Eligible proposals then go to a small multistakeholder review group convened under MAG guidance and supported by the Secretariat. The group is to assess published criteria and recommend no more than six proposals, seeking where possible an equal distribution among the three thematic streams. The MAG chooses no more than three and may take account of the balance of the portfolio across stakeholder groups, regions, gender, disciplines, technical and non-technical approaches, and developed and developing-country perspectives.

The numbers therefore describe a decision chain, not an automatic cutoff. A proposal can pass eligibility, score well on its own terms and still be displaced when the portfolio is assembled. That may be a legitimate judgement. It becomes reviewable only when the record distinguishes the proposal assessment from the later balance decision.

The process has other branches. The FAQ produced after the 3 September information webinar says similar proposals may be encouraged to merge. Public comments follow shortlisting; the PDF expressly says they are not votes and do not replace formal evaluation. The original submission, comments, full paper and revised proposal are meant to remain linked. Each branch needs lineage. A merged text is not simply the larger of two originals, and a public comment is evidence or criticism, not a mandate.

A promise about reviewers is date-sensitive

The PDF says the review-group composition will be published before the submission deadline. It also says reviewers may not submit proposals to the edition, must disclose relevant interests and must recuse themselves where a conflict exists. Those are concrete integrity commitments.

The 3 September FAQ adds uncertainty: the exact reviewer strategy depended on what arrived by the deadline and was still being finalised. The two statements can coexist. A roster can be published while assignments, expertise coverage or group structure remain unsettled. But they require a dated link between promise and implementation.

At this article's evidence cutoff, the call landing page did not name reviewers, and a bounded search of the official domain did not surface a dedicated 2026 IG Lab roster. That is not proof that none was published elsewhere. It is a verification question: where is the promised roster, which version was in force, and how will its disclosures and recusals attach to individual proposals?

Publish the boundary as a receipt

The remedy is not to forbid administrative automation. It is to preserve the chain by which administrative output can become decision input.

I propose one boundary receipt per proposal. It should begin with the submission identifier, receipt time and hash of the original version. For every AI-assisted step, it should record the permitted task, system or model version, operator, input fields, generated tag, summary, cluster or flag, and the human who accepted, changed or rejected the output.

It should then record the eligibility criterion and responsible person; any merger suggestion, comparison basis, author consent and version lineage; reviewer names, affiliations, declared interests and recusals; the assessment and shortlist recommendation; any portfolio-balance adjustment; and the MAG's final approval or rejection.

Not all of that belongs in an unrestricted public log. Contact data, protected deliberation and sensitive material can remain controlled. The public receipt can still show version, stage, responsible role, disposition and reason without exposing private inputs.

This receipt is my editorial recommendation. It is not promised by IGF or required by NIST. It applies the discipline of following the actual decision mandate rather than the institutional label, as The Policy Mirror urges. It also keeps participation distinct from authority: the Lab will stress-test proposals, but the call says it does not negotiate outcomes, establish an IGF position or endorse what it examines.

The prohibition on AI evaluation is worth preserving. A proposal-level record is how the pilot can show that the prohibition survived contact with the queue.

Sources