On Wednesday, OpenAI’s Sam Altman, Anthropic’s Dario Amodei and Hugging Face’s Clement Delangue asked the UN for common standards on AI risk, and Donald Trump’s technology adviser Michael Kratsios answered at the same conference that “International dialogue in this forum and others cannot be allowed to drift toward global governance”, the BBC reported. That reads like a dead end for multilateral AI rules. The same day, the Ada Lovelace Institute published a two-part set of reflections from six civil-society experts on what the UN’s own AI governance forum, the Global Dialogue, should deliver before it next meets, in New York in May 2027. Their priorities are practical. They ask for funded capacity, independent evaluation, fair terms for public-sector buyers and a real voice for the communities that live with AI’s effects. That is an agenda the multilateral track can work on whatever Washington thinks of it.
Strategic Insight: Kratsios told the UN that AI risks were not reason enough to constrain development with “new global governance structures”. Much of what these experts propose is capacity rather than structure: testing labs, shared contract conditions, harm-reporting channels and a funded development pot. That distinction decides how much of the UN agenda can move in the next seven months.
What did the first Global Dialogue actually produce?
The Global Dialogue first met in Geneva in July. According to the introduction to the two posts, the Heinrich Böll Stiftung brought civil-society experts from Africa, Asia and Latin America to the session, and six of them wrote forward-looking pieces that Ada published during the UN General Assembly. Ada notes that each contribution is an individual expert opinion that does not necessarily reflect the Institute’s own views.
Their verdict on Geneva is sober. Danait Girma of the Tony Blair Institute for Global Change calls the first meeting a real achievement that nonetheless set up no concrete governance mechanism. Aung Pyae of Chulalongkorn University in Bangkok says Geneva ended with an instruction, not a result: to “measure progress by actions taken, not principles adopted”. Raymond Amumpaire surveyed experts from Africa, the Middle East and Asia-Pacific on what they expected from Geneva AI Week, and concludes that those hopes were partly met at best.
We covered the run-up to Geneva in What the UN’s AI governance push means for UK business, which argued that fragmented rules are already a cost for firms trading across borders. These new pieces are more practical. They set out what should actually be done between now and May.
Why the capacity gap is the real story
The common thread is that many countries cannot check the AI systems they already depend on. Pyae cites the UN’s Independent International Scientific Panel on AI, whose preliminary report found that compute, data and evaluation expertise sit where AI is built. That leaves many countries relying on systems they cannot inspect, audit or adapt. Girma puts it more bluntly: “Some are shaping AI systems. Most are receiving them.”
Pyae also draws a distinction that matters for anyone selling AI abroad. Data sovereignty and digital sovereignty are claims over assets, and a state can hold both, he argues, without any independent way to know whether a model discriminates against its own citizens. Offers of access from the big powers do not close that gap. He cites China’s pledge of 5,000 places on AI courses and seminars for developing countries within five years, and US promotion of full-stack AI exports to allies. Using a system differs from being able to evaluate or govern it.
| Metric | Value | What it signals |
|---|---|---|
| Written submissions to the Dialogue’s preparatory process | 1,532 unique submissions (per Pyae) | Broad interest, which raises the cost of producing nothing concrete |
| Proposed Global Fund for AI target | $3 billion for skills, data and cheaper compute, proposed by the UN Secretary-General (per Girma) | The main test of whether capacity-building is funded or rhetorical |
| China’s capacity-building pledge | 5,000 places on AI courses and seminars within five years (per Pyae) | Access and training are on offer from rival powers; evaluation capacity is not |
| Time for Global Majority submissions in cluster sessions | Three minutes (per Amumpaire) | The gap between being present and having influence |
Strategic Reality: Girma argues that the UN General Assembly’s High-level Week this month should settle who runs the proposed Global Fund for AI, who pays into it and how it pays out. If it does not, the capacity agenda has a target and no money behind it.
What are the experts asking the UN to deliver?
The six contributions differ in emphasis, but they converge on four deliverables. Each is concrete enough to check in May 2027.
Independent evaluation as a named deliverable
Pyae’s first priority is to make independent evaluation a capacity-building deliverable in its own right, separate from access to technology or training. He means whether a national or regional authority can find out for itself what a system in use actually does. The outputs he proposes are benchmarks in under-resourced languages, and accredited testing laboratories whose results are mutually recognised across borders. He models this on the accreditation approach found in food safety and telecommunications. His second priority is a structured channel for regional harm evidence, much of which sits with police, fact-checkers and civil-society monitors, to reach the Scientific Panel in time for its May 2027 annual report. Otherwise, he warns, that report will mostly capture harms that are easy for European and North American institutions to see.
Critical Context: Pyae assigns his first two priorities to UN institutions and the diplomats who represent member states. His third is a job for well-resourced European and North American research bodies.
A shared baseline for buying public-sector AI
Danai Hazel Kudya of the Africa Governance and Civic Innovation Hub raises a problem any public-sector AI supplier will recognise. A government department can be legally accountable for a decision that AI helped to make whilst having little real say over how that AI system works. Governments with market power can negotiate audit rights, change notices and exit terms. Others take what they are offered.
Kudya proposes that the Dialogue convene governments, providers, standards bodies and affected communities to agree minimum operating conditions for public institutions that buy AI. Her list includes enough access to records to review decisions, notice and authorisation of material system changes, the power to restrict or suspend use, continuity arrangements if a provider has to be replaced, and a way for an individual’s challenge to fix the wider system. She suggests the Dialogue could put a short statement of these conditions before member states by 2027. “This would not create a global regulator or transfer domestic responsibility to the UN,” she writes.
A governance floor that adapts locally
Zar Motik Adisuryo of the Oxford Internet Institute uses Indonesia to argue for a global baseline of governance capabilities: minimum abilities to evaluate systems, govern data and assign accountability, which each country implements in its own way. He names two failure modes. Without a baseline, local actors fix only what is in front of them. With a rigid template, the same tools are applied to very different conditions and fail at enforcement. The aim, in his words, is “interoperability rather than uniformity”.
Participation that carries weight, and the physical costs of AI
Part 2 turns to who is heard. Amumpaire’s survey respondents described participation as tokenistic. No work between sessions included Global Majority actors, and the three minutes allotted for their submissions in cluster sessions were, he writes, “clearly insufficient”. Mathew Lubari of Community Creativity for Development describes three rooms at Geneva that he says must start talking to each other: corporate optimism, cautious diplomacy and grassroots communities living with the harms.
Lubari also widens the scope. He argues that current discussion focuses heavily on algorithmic safety and gives too little attention to AI’s physical supply chain: mining, data centres, energy use and e-waste. He wants UN institutions to consider mandatory right-to-repair principles and baseline standards across the whole AI value chain, covering labour rights, resource use and the distribution of value. Amumpaire notes that the International Labour Organisation used the Geneva discussions to promote its platform-economy decent work convention (No. 193).
- Evaluation capacity is treated as a safety issue, not a development favour. Pyae argues that governance capacity is what turns evidence of harm into protection, so safety and capacity are one priority rather than two competing ones.
- The buyer’s position is the weak point. Kudya’s case rests on the gap between legal answerability and practical control, which can be set partly by contracts and provider policies.
- Funded participation is part of the method. Pyae calls funded civil-society participation from fragile states “not a procedural courtesy”, because it keeps evaluation tied to the people it is meant to protect.
- Scope now includes hardware and labour. Lubari and Amumpaire both push the Dialogue past model risk to infrastructure and labour conditions, and Lubari adds e-waste.
⚠️ Warning: Girma warns that unless one body is named to monitor progress and report on it publicly, the Dialogue risks becoming “an annual restatement of intents rather than a mechanism for delivery”. The same risk applies to any firm that files the Dialogue under diplomacy and stops watching.
Why this matters for UK organisations
Most of these contributions draw on experience in Southeast Asia and Africa, and neither post mentions the UK. The UK relevance is indirect but real, and it comes through markets rather than law.
The most direct exposure is for UK firms selling AI into public sectors abroad. If Kudya’s baseline conditions become the questions governments routinely ask in procurement, suppliers that can already answer them will find tenders easier. Suppliers that cannot will find those questions arriving in the contract. The same logic applies to evaluation. Accredited, mutually recognised testing in under-resourced languages would create demand for evaluation skills and evidence that UK testing and assurance firms could compete to supply.
| Stakeholder group | Primary impacts | Support needs | Success metrics |
|---|---|---|---|
| UK AI vendors selling to overseas governments | Buyers may ask for record access, change notices, suspension rights and exit terms | Contract templates and technical logging that can meet those terms | Tender questions answered without bespoke engineering |
| UK AI testing and assurance providers | Possible demand for accredited, cross-border evaluation, including non-English benchmarks | Language coverage and accreditation-ready methods | Results accepted by overseas authorities |
| UK public-sector AI buyers | Kudya frames the answerability gap around all public institutions, rich country or poor | Procurement checklists covering control, not just performance | Ability to explain, pause or correct a system without the supplier’s permission |
| UK research funders and institutions | Pyae’s third priority names European and North American research bodies as partners | Partnerships with regional evaluation bodies | Funded participation that shows up in the 2027 Scientific Panel evidence |
What decides whether any of this happens
Two tests will tell. The first is money. Girma frames the proposed Global Fund for AI as the point where ambition either becomes a functioning fund, with governance, contributors and a way to pay out, or it does not. The second is ownership. Girma wants the co-chairs’ declaration turned into a programme of work, with one named body monitoring it and reporting in public. She also wants the May sessions ordered by dependency, with technical and governance capacity first, rather than run as parallel tracks.
🎯 Success Factor: Watch for named owners and dated deliverables. A Dialogue that leaves New York with an evaluation workstream, a published procurement baseline and a funded Global Fund has moved. One that leaves with a new declaration has not.
How to prepare before May 2027
💡 Implementation Framework: Tracking the UN AI agenda to May 2027
Phase 1: Map your exposure (October to December 2026)
- List contracts where an overseas public body relies on your AI system
- Check whether you could give a buyer decision records, change notice and a clean exit
- Note which of your evaluations cover languages other than English
Phase 2: Close the obvious gaps (January to March 2027)
- Draft standard contract terms for change notification and suspension
- Document how your systems are tested, in a form an outside authority could read
- Follow whether the proposed Global Fund for AI gets a disbursement mechanism
Phase 3: Read the New York outcome (April to June 2027)
- Compare the May 2027 outcome with the four deliverables above
- Check the Scientific Panel’s 2027 report for regional harm evidence
- Update procurement responses to match any published baseline
Priority actions for different organisations
If you sell AI to public bodies abroad
- Treat Kudya’s list as a tender checklist: record access, change notice, authorisation of changes, suspension and exit are reasonable things for any government to ask, whether or not the UN adopts them.
- Separate access from control in your pitch: a buyer who can use your system but cannot review or pause it carries a risk that Kudya’s list is designed to expose.
- Plan for local evaluation: assume an authority may want to test your system in its own languages and against its own cases.
If you build, test or assure AI
- Look at the accreditation model: Pyae’s proposal borrows from food safety and telecoms, where accredited lab results travel across borders.
- Invest in non-English benchmarks: under-resourced languages are named explicitly as a gap.
- Keep evidence reusable: testing documentation written once to a clear standard can serve several buyers.
If you buy AI for a UK public body
- Ask Kudya’s questions of your own suppliers: the answerability gap is not unique to the Global Majority.
- Write suspension and exit into contracts: the time to negotiate control is before deployment.
- Record who can correct the system: an affected person’s challenge should be able to fix the system as well as the individual outcome.
Resource Reality: For most firms the first step is a contract and documentation review, plus a periodic check on the Dialogue’s progress. The work grows only if a published baseline turns into a standard procurement requirement in markets you already serve.
Four things that could stall the agenda
Challenge 1: The US position hardens into non-cooperation
Kratsios warned against UN talks sliding into anything resembling global governance, and the BBC reports that he rejected any new AI regulation. If US-based suppliers treat any shared procurement baseline as governance by the back door, buyers without market power gain little from it.
Mitigation Strategy: Frame shared conditions as buyer-side procurement practice, which is how Kudya presents them, not as rules imposed on providers. Watch whether buyers adopt them jointly.
Challenge 2: Participation stays symbolic
Amumpaire’s respondents saw tokenism in Geneva, and no intersessional work included Global Majority actors. If that repeats, the evidence reaching the Scientific Panel will again reflect the harms most visible to European and North American institutions.
Mitigation Strategy: Look for funded participation and structured harm-reporting channels before May. Their absence is an early sign that the 2027 assessment will be narrow.
Challenge 3: The fund stays a target
A $3 billion proposal without contributors or a way to pay out does nothing for evaluation capacity. Girma ties the fund’s credibility to decisions in this month’s High-level Week.
Mitigation Strategy: Track concrete commitments rather than announcements. If the fund does not get a mechanism, expect the capacity agenda to shrink to training offers from individual powers.
Challenge 4: Scope creep dilutes delivery
Lubari’s case for including minerals, energy, labour and e-waste is strong on its merits, but a wider agenda is harder to deliver in seven months. The Dialogue could end up debating everything and delivering nothing.
Mitigation Strategy: Girma’s proposal to sequence sessions by dependency, with capacity first, is the practical answer. A smaller set of dated deliverables beats a comprehensive declaration.
Reality Check: Do not expect binding rules by May 2027. The realistic prize is a published procurement baseline, a named evaluation workstream and a funded Global Fund. Those would be modest wins, and they would still be more than Geneva produced.
The strategic takeaway
This week’s exchange at the UN set AI company leaders who want common standards against a US administration that will not accept global governance. The Ada-published reflections suggest that the UN track does not have to wait for that argument to be settled. Its most concrete items, including evaluation capacity, procurement conditions, harm evidence and a funded development pot, are practical infrastructure that member states and UN institutions can build.
Three factors will decide whether the Dialogue matters to UK organisations:
- Whether evaluation becomes a named deliverable: accredited, cross-border testing would change what buyers expect from suppliers.
- Whether a procurement baseline is published: Kudya’s conditions would give every public buyer the same starting questions.
- Whether the proposed Global Fund for AI is funded: without money, capacity-building stays a slogan.
Measuring progress by what gets built
Pyae’s reading of Geneva sets the right standard for May: actions, not principles. For UK firms the equivalent test is simple. Could you give an overseas public buyer the records, notice, suspension rights and exit terms that Kudya lists? Could an outside authority test your system in its own language? If yes, the Dialogue’s direction of travel is an advantage. If no, those questions will reach you through tenders whatever Washington says.
Strategic Insight: Lubari ends Part 2 with a warning that applies well beyond the Global Majority: “Until these three rooms converge, we are not truly governing AI.” For suppliers, the practical version is that the companies, diplomats and affected communities are asking different questions, and the next contract may ask all of them.
Your next steps
Immediate actions (this week):
- Identify contracts where an overseas public body depends on your AI system
- Check whether your systems keep the decision records a buyer would need for review
- Read the related coverage: AI lab chiefs ask the UN for shared risk standards; the US says no
Strategic priorities (this quarter):
- Draft standard terms for change notice, suspension and exit
- Review which evaluation evidence you hold for non-English languages
- Follow whether the proposed Global Fund for AI gets a disbursement mechanism
Long-term considerations (this year):
- Prepare for accreditation-style testing if mutual-recognition schemes emerge
- Map AI hardware repair, reuse and disposal in your supply chain
- Review the May 2027 outcome against the four deliverables above
Source and attribution
This analysis draws on “What should the UN Global Dialogue on AI Governance prioritise before its next meeting?”, a two-part series published by the Ada Lovelace Institute on 23 September 2026. Part 1 carries reflections by Danai Hazel Kudya, Aung Pyae, Danait Girma and Zar Motik Adisuryo. Part 2 carries reflections by Raymond Amumpaire and Mathew Lubari. The contributors attended the Geneva session at the invitation of the Heinrich Böll Stiftung, and their views are their own rather than the Institute’s. The remarks by Michael Kratsios are as reported by BBC News.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the proposed Global Fund for AI gets a working mechanism before the Dialogue reconvenes in New York. Read more analysis at Insights, or get in touch.