Summary

  • PNAI's second 2026 planning meeting on 20 August supported a study of AI governance beyond formal compliance but said the proposed scope remained too broad.
  • Participants asked for practical, policy-oriented and evidence-based cases, warned against a catalogue of success stories, and wanted comparable analytical criteria across jurisdictions and stakeholder groups.
  • The meeting record says the draft had not been circulated in advance, limiting detailed comment; the next step is a shared document, written feedback, revision and another meeting.
  • Daniel Kade proposes a public case-admission register showing how candidates were found, what evidence and affected voices they contain, and why each was admitted, excluded, deferred, merged or withdrawn.
  • This is an editorial proposal, not a PNAI rule. No reviewed source identifies a final theme, selected cohort, adopted method or completed 2026 output.

The sample starts before the spreadsheet

The summary of PNAI's second virtual meeting records an encouraging refusal. Participants did not want the 2026 report to become a descriptive inventory of AI governance initiatives. Nor did they want a compilation of success stories. They asked instead for implementation experience, lessons, governance outcomes and practical measures that policymakers could use in different settings.

They also saw the difficulty. “AI Governance Beyond Compliance” was supported as a direction, but it covered too much. The discussion ranged across synthetic media, misinformation, content moderation, recommender systems, chatbots, audits, transparency, accountability, interoperability and work. A report attempting all of them would either stay shallow or let the easiest examples determine its centre.

The meeting proposed comparable criteria. Cases should reveal who makes decisions, how affected communities can challenge them, and whether the arrangement supports accountability, participation, human agency and transparency. Interoperability should test whether a practice travels across legal, institutional and technical contexts. Global South experience should not arrive as decorative geography after the categories have already been fixed.

Those are useful analytical fields. They do not answer the earlier question: which cases get a row?

An open call will attract organizations able and willing to narrate their work. Successful programmes often have staff, reports and communications material ready. Failed deployments, contested decisions, informal worker experience and communities that never received a remedy are harder to document. Direct outreach can fill gaps, but it also gives the person doing the outreach power over the sample. Desk research adds independence while introducing another set of search and language biases.

If those routes are mixed without a record, the finished report can be balanced on paper and still rest on a convenience sample.

Circulate the method before asking people to trust it

The meeting summary records that the draft document had not been circulated before the call. Many participants therefore could not give detailed comments. The response was practical: circulate the work plan and conceptual document through a common collaborative file, invite written comments, revise them and hold a later meeting.

That sequence matters because scope remains reversible. Once case collection begins, an early theme becomes an intake form, the intake form favours certain kinds of evidence, and the selected examples start shaping the report's claims. Changing the method later can require discarding work and asking already burdened contributors for more material.

PNAI should attach a case-admission register to the shared method. Every submitted, solicited or discovered candidate would receive a stable reference. The public entry would record its source route, jurisdiction and development context; the submitter's role and declared relationship; the affected stakeholder position represented; the proposed theme; and the minimum evidence available for authority, challenge, implementation and observed outcome.

The disposition should be equally visible: admitted, excluded, deferred, merged, withdrawn or unresolved. A short reason and the responsible editorial decision-maker are more useful than a silent deletion. A later correction should append a state rather than erase the original choice.

Sensitive evidence can remain protected. The register need not expose a complainant, confidential audit, security detail or unverified allegation. It can say that restricted evidence exists, who holds it, what proposition it supports and what limit that places on public verification. Small cohort cells can be combined or suppressed where disclosure would identify a person.

A denominator is not a claim of representativeness

The accompanying cohort summary should count candidates received, candidates actively solicited, cases admitted and cases excluded. It should show regions, development contexts, stakeholder roles, themes, evidence classes and outcome types, along with known gaps. The purpose is not to award a diversity score. It is to stop a selected numerator from hiding the denominator that produced it.

This distinction is especially important for Global South perspectives. One case from a country cannot speak for a region. A government official, civil-society organization, local researcher, worker and affected user can see different control surfaces inside the same deployment. Their presence should be described as evidence positions, not silently upgraded into a collective mandate.

Lu Heng's Multi-Stakeholder Mirage captures the institutional danger: attendance and participation can acquire authority they were never delegated. A case register applies the same discipline to evidence. Submission gives a case standing for examination, not authority to represent everyone absent from the record.

The register will not manufacture representativeness. It will reveal whether the final language should say “the selected cases show,” “several implementers reported,” or only “one documented example suggests.” That is a meaningful improvement in the policy mirror.

PNAI decides; the IGF does not negotiate an AI rule

The current PNAI page describes an open network that shares evidence, compares approaches and co-creates guidance. Its 2026 discussions are expected to feed a new output shaped by the community. The 2025 policy brief adds an important limit: its multistakeholder recommendations were not a detailed work plan or resource commitment, and Policy Networks operated under one-year mandates without dedicated implementation budgets.

The IGF's own account says the forum does not produce negotiated outcomes. Its work can inform public and private decision-makers. It does not regulate an AI system, certify a practice or prove that a selected case is transferable.

That allocation should remain visible in the final report. Contributors provide evidence. Drafters classify and compare it. PNAI makes editorial choices about its output. Downstream institutions decide whether and how to use a recommendation under their own authority. A transparent admission register does not move those powers. It stops the narrative from implying that selection itself settled them.

Lu Heng's Policy Mirror offers the right test: a representation should remain linked to the real source, decision and correction path. For PNAI, the mirror begins at intake, not at publication.

The cost of the record is lowest now

The case-study methodology is still being narrowed. No checked public source identifies a final set of themes, selected cases or adopted indicators. That uncertainty is not a weakness to disguise; it is the reason to record the next decisions while they are cheap to reverse.

A lightweight register can begin with a dozen fields. It can later gain a correction history, missing-voice flags and a public cohort summary. What it should not do is wait until the prose is complete, then reconstruct why a vivid success was retained while a difficult counterexample disappeared.

PNAI has already stated the right ambition: move beyond compliance and beyond catalogues. The next proof is procedural. Before telling readers what the cases mean, show them how the cases became the cases.

Sources