Fathom

The Global Dialogue's distinct value lies in moving beyond principles toward the practical infrastructure that makes AI governance work. Member States and stakeholders have already articulated shared values across the GDC, the UNESCO Ethics Recommendation, the OECD Principles, and regional frameworks. The persistent gap is implementation: how those values are operationalized, verified, and held accountable in practice. A successful first Dialogue would advance three concrete outcomes. First, it would establish that independent, expert-led verification—alongside, not instead of, traditional governmental oversight—is a core element of credible AI governance. Just as financial markets rely on independent auditors and physical product safety relies on accredited conformity assessment bodies, AI governance requires technical assessment capacity that is independent of both developers and political cycles. Recognizing this category in the Dialogue's vocabulary clarifies and paves the way for effective governance. Second, it would identify where international coordination on verification standards adds the most value: shared definitions of unacceptable risks, mutual recognition mechanisms so that verification in one jurisdiction can be relied upon in another, and pooled technical and legal capacity for jurisdictions that cannot stand up assessment infrastructure alone. Third, it would create a durable structure for ongoing work between annual convenings, including channels through which national and subnational governance experiments—from the EU AI Act's conformity assessment regime to emerging state-level frameworks in the United States to AI Safety Institutes—can inform one another. The measure of success is not consensus on a single global rulebook, which is neither achievable nor desirable on this timeline. It is whether the Dialogue produces the connective tissue—vocabulary, mutual recognition, shared infrastructure—that allows diverse governance approaches to function as a coherent system rather than a fragmented patchwork.

• AI capacity-building
• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

These four priorities cohere around a single proposition: trustworthy AI governance requires verification infrastructure-the institutional capacity to assess, in technical detail, whether AI systems meet publicly-set standards of safety and accountability. Safe, secure and trustworthy AI. Safety claims that cannot be independently verified are commitments without consequence, or worse, safety theater. Establishing shared expectations for what verification looks like-who conducts it, against what standards, with what transparency-is foundational to every other governance objective. AI capacity-building. Verification capacity is unevenly distributed globally. Many jurisdictions face a translation problem rather than a will problem: clear policy intent, but no domestic infrastructure to assess whether AI systems comply. Building shared technical and accreditation infrastructure that can serve multiple jurisdictions-rather than each building its own from scratch-is among the most consequential capacity-building interventions the Dialogue could advance. Interoperability of governance approaches. Governance regimes are proliferating. Without mutual recognition mechanisms, companies face redundant compliance burdens and jurisdictions cannot rely on one another's assessments. Interoperability does not require harmonized rules; it requires shared standards and approaches for the verification process itself, so that determinations made under one regime can be understood and, where appropriate, accepted under another. Legibility is key. Transparency, accountability, and human oversight. Independent verification operationalizes these principles. Transparency requires that assessment criteria, methodologies, and findings be publicly accessible. Accountability requires that verification bodies themselves be subject to oversight, with revocable accreditation and safeguards against conflicts of interest and race to the bottom dynamics. Human oversight requires institutions-not only individuals-empowered to make consequential judgments about AI systems on the public's behalf. Together, these priorities point toward verification as the practical mechanism by which abstract governance principles become enforceable practice.

Several cross-cutting issues warrant explicit attention. The accreditation gap. Across jurisdictions, governance frameworks increasingly rely on third-party assessments-conformity assessments under the EU AI Act, audits in sectoral regulations, verification regimes in emerging national and subnational laws. Yet there is no shared international understanding of what qualifies an organization to perform these assessments competently and independently. Without accreditation infrastructure, third-party assessment becomes a market for credentialing without rigor, and verification claims become unreliable signals. This is a problem the international community has solved before, in domains from medical devices to financial auditing to electrical safety, through accreditation bodies that operate at arm's length from both industry and political authorities. AI requires comparable infrastructure, and the Dialogue is uniquely positioned to identify principles for it. Mutual recognition and the cost of fragmentation. As verification regimes emerge, the question of whether assessments conducted in one jurisdiction can be relied upon in another will determine whether the global system is coherent or fragmented. Fragmentation imposes the heaviest costs on smaller economies and developing-country regulators, who face the choice between accepting assessments designed for other contexts or building parallel infrastructure they cannot afford. The Dialogue can begin establishing the principles-shared accreditation criteria, transparent methodologies, recognized standards bodies-that make mutual recognition possible without requiring rule harmonization.

Across jurisdictions, governments are demonstrating a similar pattern: legislative momentum to act on AI risks, followed by stalled implementation when the technical work of standard-setting and assessment proves harder than the legislation anticipated. The implementation bottleneck. Laws and frameworks increasingly require that AI systems meet specified standards—for safety, non-discrimination, transparency, child protection, or critical-infrastructure security—and that compliance be assessed by qualified parties. But the qualified parties themselves often do not yet exist in adequate number or with credible accreditation. The result is laws on the books that cannot be fully enforced and businesses uncertain about how to demonstrate compliance, assurance, and trust. Further, AI legislation has proven to be procedural rather than outcome-based in nature and have not been able to keep up with the pace of technology. A more agile system that catalyzes the field of outcomes-based evaluations is needed. Asymmetric capacity. This problem is most acute for jurisdictions outside the small group of countries with deep domestic AI assessment expertise. Without shared infrastructure, every jurisdiction is asked to build its own verification capacity from scratch—duplicative for well-resourced states and prohibitive for many others. The likely outcome, absent coordination, is either dependence on foreign assessment regimes designed around foreign priorities, or under-enforcement of domestic law. Specific harms in the meantime. While verification infrastructure is being built, harms continue to accrue, particularly in domains where AI systems interact directly with vulnerable populations—children using conversational AI products, individuals affected by AI-mediated decisions in employment, credit, and public services, and communities exposed to AI-enabled fraud and disinformation. The opportunity. These challenges are tractable. The international community has built accreditation systems for other technical domains, and the AI assurance ecosystem—research labs, audit firms, civil society technical organizations—is growing rapidly. The gap is the connective tissue between policy intent, technical capacity, and accountable institutions. Closing it is the work that international cooperation is best suited to support.

The Dialogue can play a role no other forum is positioned to play by legitimizing and connecting the practical infrastructure that makes AI governance work across jurisdictions. Establishing shared vocabulary. International cooperation requires common terms. Concepts such as independent verification, accreditation, conformity assessment, and mutual recognition are well-developed in other regulated domains but are still being adapted for AI. The Dialogue can give these concepts a common reference point, so that jurisdictions designing oversight regimes are working from a shared conceptual foundation rather than reinventing it. Facilitating mutual recognition. Where jurisdictions converge on comparable verification standards, the Dialogue can support the diplomatic and technical work of recognizing one another's assessments. This reduces compliance costs, strengthens the value of verification, and prevents AI assurance from becoming a fragmented patchwork in which the same system must be reassessed in each jurisdiction it enters. Supporting capacity-sharing. Jurisdictions without domestic verification infrastructure should not be forced to choose between under-enforcement and dependence. The Dialogue can convene the work of building shared accreditation bodies, pooled technical capacity, and reference standards that any participating jurisdiction can adopt or adapt—analogous to how international standards bodies serve countries that could not produce equivalent standards alone. Coordinating crisis response. When AI incidents have cross-border consequences, the Dialogue can serve as the standing channel through which Member States communicate, share verified information, and coordinate responses. Establishing such a channel before it is urgently needed is itself a form of preventive governance. The Dialogue's comparative advantage is not in producing binding rules quickly, but in producing the durable infrastructure that allows national rules—wherever they are made—to function as part of a coherent global system that broadly advances safety and security.

The Dialogue should connect with the following ecosystems: National and regional verification regimes. The EU AI Act's conformity assessment system; emerging US state frameworks authorizing independent verification organizations, including bipartisan legislation in California, Connecticut, Ohio;, and Virginia; the UK's AI Security Institute and the international Network of AI Safety Institutes; sectoral assessment regimes in finance, health, and critical infrastructure across multiple jurisdictions. Accreditation and standards infrastructure. ISO/IEC standards on AI management systems, risk management, and trustworthiness; IEEE technical standards; ITU recommendations; the International Accreditation Forum and national accreditation bodies that operate the global mutual-recognition system for conformity assessment in adjacent domains. Our organization, Fathom, is designing a systems-wide assessment for AI. Scientific and research bodies. The Independent International Scientific Panel on AI; the International AI Safety Report; the Singapore Consensus on AI Safety Research Priorities; OECD AI Policy Observatory and Global Partnership on AI; civil society research institutions developing technical evaluations. Multilateral governance frameworks. UNESCO Recommendation on the Ethics of AI; the Hiroshima AI Process; AU Continental AI Strategy; ASEAN Guide on AI Governance and Ethics; Council of Europe Framework Convention on AI. The Dialogue's added value. The added value the Dialogue can offer to this ecosystem is connective rather than substitutive. Each of the initiatives above is producing valuable work in its domain. What is missing is the standing forum where they can be brought into relation with one another, where gaps and overlaps can be surfaced, where small and developing-country jurisdictions can engage on equal footing, and where the legitimacy of the universal multilateral system can support the broader goal of interoperable governance. The Dialogue should treat its role as catalytic and connective, not duplicative.

Effective participation requires structures that engage stakeholders' expertise rather than treating them as audiences for governmental conclusions. Standing technical working groups. Annual convenings alone cannot carry forward technical work on verification standards, accreditation principles, or mutual recognition. The Dialogue should establish thematic working groups that operate continuously between annual sessions, with formal channels for civil society technical organizations, accreditation bodies, AI Safety Institutes, and assurance practitioners to contribute substantive proposals—not merely commentary on government drafts. A consortium model for capacity-sharing. On topics requiring sustained technical infrastructure—accreditation, shared evaluation methodologies, model verification criteria—the Dialogue should support consortium structures in which jurisdictions join together to build and govern shared infrastructure. This pattern works in adjacent domains and avoids duplicative investment. Genuine inclusion of affected communities. Communities directly affected by AI harms—children and parents navigating generative AI products, workers subject to algorithmic management, communities affected by AI-mediated public-services decisions—bring knowledge that technical experts cannot substitute. Their participation should be structured through ongoing channels in multiple languages and formats, with documented evidence of how their input shapes outcomes. Meaningful inclusion of subnational and regional governance. Significant AI governance experimentation is happening at subnational levels—US states, EU member states implementing the AI Act, provincial and devolved authorities in many federal systems. Excluding this layer would forfeit substantial practical learning. The Dialogue should create formal channels through which subnational governance lessons can be surfaced to national delegations and to other jurisdictions facing similar design choices. Transparency in process. Public agendas, open submissions, plain-language summaries of decisions, and documented responses to stakeholder input are practical requirements for legitimacy.

Several categories of underrepresentation are persistent. Communities directly affected by AI harms. Children and the parents and educators responsible for them; workers subject to algorithmic management and AI-driven displacement; individuals whose access to credit, employment, housing, or public services is mediated by AI systems; communities targeted by AI-enabled fraud, harassment, or disinformation. These groups hold knowledge about system failures that technical analysis cannot reproduce. Inclusion requires sustained channels—in multiple languages and accessible formats—rather than one-off consultations. Subnational and regional governance actors. In federal systems and devolved governance regimes, much of the practical work of AI oversight occurs below the national level. State, provincial, and regional officials designing and implementing AI governance frameworks bring directly relevant operational expertise that national delegations may not fully capture. Formal channels for their participation strengthen the Dialogue's connection to implementation realities. Smaller and developing-country regulators. Regulators in jurisdictions without large domestic AI sectors face distinctive challenges and design constraints. Their underrepresentation skews global discussions toward the priorities of producing-country regulators. Capacity support for sustained participation—including technical staff support, translation, and travel where needed—is a practical requirement of equitable inclusion. The technical assurance community. Independent auditors, evaluation researchers, conformity assessment practitioners, and accreditation specialists possess essential implementation knowledge that is often missing from AI policy discussions. Their inclusion through dedicated technical channels would substantially strengthen the Dialogue's substantive output. The judicial and legal sectors. Judges, legal scholars, and court systems will increasingly adjudicate disputes involving AI systems. Their perspective on evidence, liability, and remedies should inform the design of governance frameworks, not be left to retrofit afterward.

Formats should be chosen for substantive yield and could include: Live technical demonstrations and assessments. Rather than abstract discussion of AI capabilities and risks, the Dialogue could host structured sessions in which independent technical experts demonstrate specific capabilities, evaluation methodologies, and verification techniques in real time. Participants—including non-technical policymakers—would observe what an evaluation actually looks like, what it can and cannot determine, and how findings translate into governance decisions. This grounds discussion in technical reality, connected with policy reality. Cross-jurisdictional implementation roundtables. Convenings organized around specific governance challenges—verifying AI safety in products used by minors, assessing AI systems deployed in critical infrastructure, evaluating AI in government services—that bring together regulators, technical experts, affected communities, and assessment practitioners from multiple jurisdictions. The format should privilege specific cases over general principles, surfacing concrete trade-offs and design choices. Standing working sessions with documented outputs. Working groups whose deliberations are open, whose proposals are published in draft for consultation, and whose final recommendations are formally received by the Dialogue. This produces a public record of substantive work between annual convenings and creates accountability for follow-through. Affected-community testimony with structured response. Testimony from communities directly affected by AI harms is most useful when paired with structured response: a designated technical respondent, a regulator respondent, and a documented outcome describing what action, if any, the testimony shaped. Without this structure, testimony risks becoming ceremonial. Capacity-building convenings. Targeted sessions in which jurisdictions building verification or oversight infrastructure can engage directly with peers further along in implementation, with technical support from accreditation and standards bodies. These should be designed as working sessions, not showcases.

Several emerging approaches illustrate how independent verification can operationalize governance objectives across diverse legal traditions. Independent Verification Organizations (United States, multi-state). A bipartisan model in which state legislatures authorize accredited, independent expert bodies to verify AI systems against publicly-set safety outcomes. States define the goals; licensed verification organizations translate them into technical criteria and assess submitting developers. Verification is opt-in but carries meaningful incentives, and accreditation is performance-based and revocable. Legislation has been passed in Virginia and Connecticut, with additional states active, and the model is supported by a politically diverse coalition of scholars including Geoffrey Hinton, Yoshua Bengio, Stuart Russell, Gillian Hadfield, Jack Clark, and Anthony Aguirre. A multi-jurisdictional Center of Excellence is being organized to support interoperability and shared technical capacity. Conformity assessment under the EU AI Act. A binding regime in which high-risk AI systems must undergo conformity assessment-in some cases by notified third-party bodies-before market placement. Demonstrates how the well-established European conformity assessment infrastructure, developed across decades for products from medical devices to machinery, can be extended to AI. AI Safety Institutes. Government-affiliated technical bodies in the UK, US, Japan, Singapore, and Canada conducting pre-deployment evaluations of frontier AI systems and coordinating internationally. Demonstrates how technical assessment capacity can be built within governments while remaining methodologically independent. Accreditation infrastructure in adjacent domains. UL Solutions in electrical safety; the International Accreditation Forum in conformity assessment; the PCAOB in financial auditing. Each illustrates how independent technical assessment, operating under public oversight with revocable credentials, can scale across jurisdictions. AI governance can adapt these precedents rather than invent the institutional form anew. A common thread. Across these approaches, the consistent insight is that credible governance requires institutions-independent, accountable, technically expert-that can render verifiable judgments about AI systems on the public's behalf.