Submission to the UN Global Dialogue on AI Governance

Text edited for legibility.

The first session of the Global Dialogue on AI Governance^[1] arrives at a moment when AI’s trajectory is being set faster than the institutions meant to govern it can adapt.  Decisions taken in the next few years about capability, deployment, oversight and access will shape how billions of people work and think for decades. Few of those decisions are currently  subject to democratic deliberation at the international level and the Dialogue presents a unique universal forum to change that.

CFG works on exactly what the Dialogue was created to address, which is why we are submitting this contribution. Our Advanced AI programme helps policymakers calibrate how much autonomy AI systems assume so AI can be deployed with confidence while keeping human decision-making where it matters. The seven thematic areas of Resolution A/RES/79/325 map onto the questions our work is built around: where on the spectrum from human autonomy to AI autonomy a given system sits, whether governance is calibrated to that position, and who decides. The responses below offer methodology and concrete mechanisms drawn from our work.

Productive outcomes from the first Global Dialogue on AI Governance

“In your opinion, what outcomes would make the first Global Dialogue on AI Governance a success? (Max. 300 words)”

A successful first Dialogue must deliver a shared problem statement, a working methodology and a follow-up architecture.

On the problem statement: the governance struggle runs on two fronts. The first is the tension between tech companies and nation states. Firms shape finance, markets and public discourse, and enforcement authorities are catching up. The second is technology as an instrument of US–China competition, most visibly in the contest over access to and control of global data. Both share one diagnostic: the ungoverned transfer of autonomy from humans to systems whose behaviour is opaque, unaccountable and difficult to reverse. That formulation lets the seven thematic areas in Resolution A/RES/79/325 cohere around a single question: at what level of autonomy is each system operating relative to its users, and is that level governed?

On methodology: autonomy-gap analysis addresses tension between tech companies and nation states. The gap between the autonomy of law sanctions and the autonomy a system exercises in deployment is measurable. CFG’s Enforcement Tracker shows this for the EU AI Act, the Digital Services Act and the Digital Fairness Act. This methodology can transfer to international frameworks. Adoption at the UN level lets governments report on enforcement reality, not just legislative intent.

On follow-up: the co-chair summary should commit to three concrete mechanisms before the May 2027 New York session: 

1. An intergovernmental expert group on common audit methodologies across cognitive, security and labour domain;
2.  A reporting line into OHCHR, ILO and UNESCO under their existing mandates. The Dialogue should keep open alternative pathways based on UN General Assembly resolutions, where the rest of the international community can move without being blocked;
3. A public registry of national autonomy-gap findings.

Thematic areas for the AI Dialogue most relevant to CFG’s work

“From your perspective, which of the following thematic areas identified by the General Assembly Resolution 79/325 for the AI Dialogue reflect your priorities for urgent action and active engagement by your entity? Please select up to 4 priorities, and briefly explain your selection. (Max. 300 words)”

Safe, secure and trustworthy AI; Interoperability of governance approaches; Protection and promotion of human rights; Transparency, accountability, and human oversight. Three of the four are priorities for CFG. Safe, secure and trustworthy AI is the umbrella above the other three.

Transparency, accountability and human oversight is the first priority. CFG’s Enforcement Tracker measures the gap between the level of autonomy a law sanctions and the level a system exercises in deployment. We have built this evidence base for the EU AI Act, the Digital Services Act and the Digital Fairness Act. The same methodology transfers to other jurisdictions and to international monitoring. Enforcement data is critical to help governments understand their operational reality.

The right most directly at stake is freedom of thought. AI systems are already intervening in the cognitive processes that freedom of thought exists to protect. CFG proposes cognitive integrity as the governance standard for AI: the condition under which people think, decide and act based on their goals and values rather than being subverted by external systems. That standard cuts across education, healthcare, public administration and the labour market. The Dialogue should engage with cognitive integrity as a shared governance concept and bring these agencies into the conversation together.

Interoperability of governance approaches is the third priority. As Europe’s AI Act begins to take effect, deployment shifts to less regulated jurisdictions. The Dialogue’s value is its universal mandate: it is the only forum in which a common audit methodology can be developed across the cognitive, security and labour domains where autonomy transfer is most consequential. Interoperability should focus on developing shared methods for measuring whether rules are working.

Other cross-cutting or emerging issues

“In your opinion, are there any cross-cutting or emerging issues not captured by the listed themes above? If so, please explain. (Max. 300 words)”

The seven themes ask what AI does to people now. They do not ask what irreversible commitments today’s governance choices impose on people who do not yet have a vote. At CFG we are particularly interested in integer rational fairness. Autonomy transfer compounds: a system permitted to operate at a given level today shapes the workflows and institutional muscle memory that constrain what is reversible tomorrow. The Dialogue should commission a working note on intergenerational reversibility.

Compute and energy are another cross-cutting issue. A handful of companies hold the choke points: TSMC fabricates most of the advanced-node logic AI runs on, ASML is the only supplier of EUV lithography, and NVIDIA controls most of the AI accelerator market. They do not answer to the state that uses a model nor the state that trains it, so rules written at the application layer cannot touch what they decide. The Dialogue should treat compute concentration as a governance object, not something it leaves to the market.

Defence and security implications of AI systems is where autonomy transfer is happening fastest, and where the EU AI Act, the AI Liability Directive and most national frameworks do not reach. NATO, the OSCE and the UN First Committee have parallel conversations the Dialogue does not currently connect. A reporting line into the Office for Disarmament Affairs, alongside OHCHR, ILO and UNESCO, would close the gap between civilian governance and the domain where ungoverned autonomy carries the highest immediate risk.

Another important issue is the AI’s energy and water budgets. CFG’s Tech and the Planet project frames the constraint: progress within limits.

Lastly, cultural convergence is emerging as a particularly salient cross-cutting challenge. Systems trained on similar datasets and designed by similar people narrow what gets reproduced at scale. Cultural diversity is a governance question.

The role of the AI Dialogue in advancing international cooperation on AI governance

“What role can the AI Dialogue play in advancing international cooperation on AI governance? (Max. 300 words)”

The Dialogue is the only universal forum where international cooperation on AI governance can be built. The Dialogue can produce a shared diagnostic method governments can use across jurisdictions. CFG built an autonomy-gap analysis method for the EU’s digital rulebook: measuring the gap between the autonomy a law sanctions and the autonomy a system exercises in deployment. The Dialogue can host the connective architecture that civilian-coded governance currently lacks. AI’s hardest governance problems (military, compute, energy, cognition) cut across mandates no single agency holds. A reporting line from the Dialogue into OHCHR, ILO, UNESCO and the Office for Disarmament Affairs would let those bodies report on AI together rather than separately.

The Dialogue also offers legitimacy for middle-power coordination. The great power competition will not produce universal rules in the time available. Resolutions modelled on the General Assembly format let the rest of the international community move on shared standards without being blocked by, or caught between great powers. As every state can take part in the Dialogue.

The Dialogue should produce a single global regulatory text. Harmonisation of every rule is neither realistic nor desirable as different jurisdictions face different threats. They need a common method for measuring whether their rules are working, a shared vocabulary for naming the autonomy transfers their citizens face, and a calendar that holds governments to follow-up.

Existing initiatives, partnerships, or mechanisms

“What are some of the existing initiatives, partnerships, or mechanisms that the AI Dialogue should build upon or connect with, and what added value could the AI Dialogue bring? (Max. 300 words)”

The Council of Europe Framework Convention on AI is the first binding international treaty on AI. It opened for signature in September 2024 with non-Council parties already among the signatories. The rights-based architecture is the floor and the Dialogue’s job is to extend it geographically.

The OECD AI Principles and the Global Partnership on AI hold the most developed methodology base outside the EU. The Dialogue should scale that methodology to non-OECD jurisdictions and integrate it with the autonomy-gap method CFG has built for the EU AI Act, the Digital Services Act and the Digital Fairness Act. CFG’s Enforcement Tracker provides that operational data.

UNESCO’s Recommendation on the Ethics of AI is in implementation across 194 member states. UNESCO has normative reach the Dialogue should reinforce by tying enforcement reporting to it.

The AI Safety Institutes network (UK, US, Japan, Singapore and others) holds operational evaluation capacity the UN system lacks. Linking AISI protocols to UN reporting closes the methodology-to-monitoring gap. Hardware-enabled mechanisms, the chip-level governance work CFG and others have developed, are the technical infrastructure that makes verification credible at international scale, similar to IAEA inspection regimes.

The International Association for Safe and Ethical AI (IASEAI) brings the AI safety and ethics communities together. Its inaugural Paris conference in February 2025 issued a ten-point Call to Action on binding safety standards and global cooperation. The Dialogue needs that civic track alongside state-led tracks.

The High-Level Advisory Body on AI and the Independent International Scientific Panel created under A/RES/79/325 are the Dialogue’s institutional companions. They should be commissioned to produce annual autonomy-gap findings from national reporting.

Formats and structures of stakeholder contributions

“How can different stakeholders contribute to the AI Dialogue? Please share recommendations for the format and structure of the AI Dialogue. (Max. 300 words)”

Member states bring sovereignty and reporting obligations. The Dialogue should require annual national autonomy-gap reports, on the model of the Human Rights Council’s Universal Periodic Review (UPR).

Civil society and research organisations bring methodology and evidence. CFG’s Enforcement Tracker is one example: an operational dataset on EU digital rulebook enforcement. The Dialogue should build a permanent civil society track with rapporteur status across the seven themes.

Industry brings capability, deployment data and risk visibility no government has. The Dialogue should require frontier developers to submit autonomy and capability data, like financial reporting. Voluntary commitments are not enough.

Standards bodies (ISO/IEC JTC 1/SC 42, IEEE) bring technical interoperability. They should sit in working groups, not in observer status.

Parliamentarians bring democratic accountability the executive track misses. An inter-parliamentary line on the IPU–UN model would close that gap.

The format should follow three rules:

• Working tracks should be split by theme, each producing one deliverable per Dialogue cycle.
• A permanent secretariat with technical capacity is needed, as opposed to a rotating Chair.
• A public follow-up dashboard should track commitments against outcomes between sessions.

The single most useful structural change will be to tie the Dialogue’s calendar to a public reporting cycle to ensure the Dialogue is an accountability mechanism.

Underrepresented voices, communities, and perspectives in AI governance

“Which voices, communities, or perspectives are currently underrepresented in global discussions on AI governance? How could they be included? (Max. 300 words)”

Future generations cannot represent themselves because they do not yet exist. Yet today’s autonomy-transfer decisions impose constraints on people who will live with them. Wales has a Future Generations Commissioner under its 2015 Well-being of Future Generations Act and this can serve as a model. The Dialogue should require an intergenerational impact assessment for each thematic working group’s output, and create a standing intergenerational rapporteur. CFG was created to put future generations into present-day technology decisions. The Dialogue is the natural multilateral home for that work.

AI training, deployment and benefits sit in a small number of jurisdictions. Governance discussions sit in a smaller set still. The Dialogue’s format already has universal participation; what it does not yet have is universal preparation. Regional pre-Dialogue consultations, hosted by African Union, ASEAN, CARICOM, CELAC and Pacific Islands Forum secretariats and resourced from the UN budget rather than donors, would let Southern positions enter the Geneva and New York sessions on equal footing. Travel and translation budgets should not be optional.

Lastly, workers and the people whose cognition AI now mediates: pupils in classrooms, patients in clinics, citizens in administrative systems. These are the populations the seven themes are about, yet they have no seat. The Dialogue should establish a community-witness track with structured testimony from teachers, doctors, social workers and trade unions. ILO has the convening capacity for the labour side. UNESCO and WHO can convene the cognitive side.

Recommendations for innovative engagement formats

“What innovative engagement formats could most effectively foster meaningful and dynamic engagement during the AI Dialogue? (Max. 300 words)”

1. Live autonomy-gap audits. Demonstrations run in plenary using real systems and real legislation. Member states and civil society watch the gap measured in real time.
2. Citizens’ assemblies on AI governance, randomly selected and convened in parallel with the Dialogue. Ireland’s constitutional convention model is the precedent. The output is published alongside member-state submissions and tabled in working tracks.
3. AI-assisted deliberation platforms modelled on Taiwan’s vTaiwan/Polis system. The Dialogue should run them in the months between sessions to surface where actual agreement exists across jurisdictions.
4. Working-track residencies. Civil society researchers, technical experts and parliamentarians in residence with the Dialogue secretariat for six-month rotations. They produce track outputs, not observer reports.
5. Witness sessions with structured testimony from teachers, doctors, social workers, trade union representatives and frontline communities. Brief panels, public record, formal response from each working track within thirty days.

These formats produce a public artefact that can be used as a public registry of every commitment made at the Dialogue, with status updates published quarterly providing accountability tracking.

Policies, practices, platforms, or approaches to promote effective AI governance

“Please share examples of policies, practices, platforms, or approaches that promote effective AI governance or offer concrete solutions to addressing its challenges. (Max. 300 words)”

CFG’s Enforcement Tracker measures what governance actually achieves in deployment. It logs into a dataset every public enforcement action under the EU AI Act, the Digital Services Act, the Digital Markets Act and GDPR. The Tracker also enables cross-legislative learning. By comparing enforcement dynamics across different regulatory frameworks, policymakers can identify which governance approaches are more effective, where coordination gaps exist, and how lessons from one domain can inform another. Other jurisdictions can build the same monitoring against their own rulebooks. Without a tracker, regulators cannot know what their own laws are doing.

Hardware-enabled mechanisms (HEMs) are chip-level verification embedded in AI accelerators. CFG and others have proposed them as the technical infrastructure for international compute governance, analogous to IAEA inspection regimes. In practice this means tamper-evident telemetry on training runs that can be exported to neutral verifiers.

AI Safety Institute evaluations are the closest existing model for pre-deployment testing. The UK AISI evaluates frontier models pre-deployment under agreements with leading developers. The Dialogue should connect AISI protocols to UN reporting under A/RES/79/325.

Public algorithm registries are working at the city level. Amsterdam, Helsinki and others maintain registers of public-sector AI systems with risk scores and contact lines. The EU AI Act’s high-risk database extends the model. The registry is a low-cost transparency mechanism that scales to national and international levels.The Council of Europe Framework Convention on AI is the only binding international treaty with a rights-based floor. Its parties already extend beyond the Council. The Dialogue should treat the Framework Convention as the rights baseline and build interoperability with it.

“What innovative engagement formats could most effectively foster meaningful and dynamic engagement during the AI Dialogue? (Max. 300 words)”

1. Live autonomy-gap audits. Demonstrations run in plenary using real systems and real legislation. Member states and civil society watch the gap measured in real time.
2. Citizens’ assemblies on AI governance, randomly selected and convened in parallel with the Dialogue. Ireland’s constitutional convention model is the precedent. The output is not advisory; it is published alongside member-state submissions and tabled in working tracks.
3. AI-assisted deliberation platforms modelled on Taiwan’s vTaiwan/Polis system. The Dialogue should run them in the months between sessions to surface where actual agreement exists across jurisdictions.
4. Working-track residencies. Civil society researchers, technical experts and parliamentarians in residence with the Dialogue secretariat for six-month rotations. They produce track outputs, not observer reports.
5. Witness sessions with structured testimony from teachers, doctors, social workers, trade union representatives and frontline communities. Brief panels, public record, formal response from each working track within thirty days.

What unifies these formats: each produces a public artefact that can be cited the next year. The Dialogue’s risk is becoming a forum where things are said and nothing is recorded. The least innovative format is also the most necessary: a public registry of every commitment made at the Dialogue, with status updates published quarterly. Without that, even the best engagement format produces no accountability.

Policies, practices, platforms, or approaches to promote effective AI governance

“Please share examples of policies, practices, platforms, or approaches that promote effective AI governance or offer concrete solutions to addressing its challenges. (Max. 300 words)”

CFG’s Enforcement Tracker measures what governance actually achieves in deployment. It logs every public enforcement action under the EU AI Act, the Digital Services Act, the Digital Markets Act and GDPR. The Tracker turns enforcement from anecdote into dataset. The Tracker also enables cross-legislative learning. By comparing enforcement dynamics across different regulatory frameworks, policymakers can identify which governance approaches are more effective, where coordination gaps exist, and how lessons from one domain can inform another. Other jurisdictions can build the same monitoring against their own rulebooks. Without a tracker, regulators cannot know what their own laws are doing.

Hardware-enabled mechanisms (HEMs) are chip-level verification embedded in AI accelerators. CFG and others have proposed them as the technical infrastructure for international compute governance, analogous to IAEA inspection regimes. The mechanism is concrete: tamper-evident telemetry on training runs, exportable to neutral verifiers. HEMs make compute concentration governable.

AI Safety Institute evaluations are the closest existing model for pre-deployment testing. The UK AISI evaluates frontier models pre-deployment under agreements with leading developers. The Dialogue should connect AISI protocols to UN reporting under A/RES/79/325.

Public algorithm registries are working at the city level. Amsterdam, Helsinki and others maintain registers of public-sector AI systems with risk scores and contact lines. The EU AI Act’s high-risk database extends the model. The registry is a low-cost transparency mechanism that scales to national and international levels.

The Council of Europe Framework Convention on AI is the only binding international treaty with a rights-based floor. Its parties already extend beyond the Council. The Dialogue should treat the Framework Convention as the rights baseline and build interoperability with it, not duplicate it.