Signatu

First, agreement that AI governance requires formal rule infrastructure — not just policy intent. Current governance frameworks, including the EU AI Act and national AI strategies, correctly identify risks but stop short of specifying how AI systems should be made to reason correctly over legal and regulatory rules. The Dialogue should establish consensus that the gap between a governance principle and an AI system that actually enforces it must be closed with formal logic — not with guidelines, disclaimers, or audits after the fact. Rules must be machine-executable, not merely human-readable. Second, recognition of the legal hallucination crisis as a systemic governance failure, not an edge case. AI systems are already fabricating court judgments, inventing statutory provisions, and generating legally incorrect outputs at scale across jurisdictions. This is not a product liability question for individual vendors — it is a governance failure that undermines the rule of law itself. The Dialogue should name this explicitly and call for interoperable formal rule standards that any AI system operating in legal, regulatory, or rights-bearing contexts must satisfy. Third, a framework for AI agent authorization grounded in rights and obligations — not just safety ratings. As AI agents begin to act autonomously on behalf of individuals and institutions, the question is no longer whether AI is safe in the abstract, but whether it is formally authorized: what it may do, for whom, under which conditions, and with what legal basis. The Dialogue should establish that agent authorization requires a formally complete rule layer — one that encodes rights, obligations, permissions, and prohibitions with the same rigour that law demands of human actors. Governance without formal rule infrastructure is aspiration. The Dialogue succeeds if it begins closing that gap.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches
• Protection and promotion of human rights

Our work on KROG Rules - a formally complete rule system for AI reasoning over law, contracts, and regulations - maps directly to these four priorities. Safe, secure and trustworthy AI is our core technical contribution. AI systems operating in legal and regulatory contexts currently fabricate court judgments, invent statutory provisions, and hallucinate contract clauses because they have no formal logic layer. KROG provides that layer: a mathematically complete system of 7 foundation types, 35 relationship rules, and 9 modal operators that encodes any law, contract, or regulation so that AI reasons over it correctly. Trustworthy AI in high-stakes domains requires formal rule infrastructure - not just better language models. Transparency, accountability, and human oversight becomes technically achievable when AI outputs are grounded in formal rules. KROG makes every AI decision traceable to the specific rule it was checked against. This is not transparency as a reporting obligation - it is transparency as a structural property of the system. Oversight requires that humans can verify AI reasoning, and verification requires that reasoning is formally encoded. Protection and promotion of human rights connects directly to our work on personal data wallets under the EU Digital Identity framework. Rights such as data access, erasure, and portability exist in law but cannot be exercised without tooling that encodes them as machine-executable obligations. KROG provides the rule engine that translates legal rights into enforceable digital actions - rights that cannot be technically enforced remain aspirational. Interoperability of governance approaches is addressed by KROG's structural property of cross-domain isomorphism: the same formal rule language governs contracts, regulations, and rights obligations across jurisdictions. A shared machine-readable rule layer enables different national governance frameworks to become interoperable without requiring political harmonisation - technical interoperability can precede and enable regulatory convergence. Governance without formal rule infrastructure is intention. These four priorities are where that gap closes.

Two cross-cutting issues are absent from the listed themes and deserve explicit recognition. The AI legal reasoning crisis as a systemic rule-of-law threat. The listed themes address AI safety, rights, and accountability in general terms. None captures a specific and already-occurring failure: AI systems are fabricating court judgments that do not exist, inventing statutory provisions, and generating legally incorrect outputs that are entering judicial proceedings across jurisdictions. A tracker of such cases identifies over 600 documented instances in US courts alone, with the rate accelerating to multiple new cases per day as of 2025. This is not a theoretical risk - it is an active degradation of legal system integrity. The Dialogue should name the formal correctness of AI reasoning over law as a distinct governance requirement, separate from general AI safety. An AI system that is safe in the abstract but legally incorrect in practice undermines the rule of law in ways that no safety rating captures. Formal authorization infrastructure for AI agents. The listed themes address human oversight of AI, but none addresses the structural question that agentic AI makes urgent: when an AI system acts autonomously on behalf of a person or institution - making decisions, entering obligations, executing transactions - what formal rule system defines what it is authorized to do, for whom, and under which conditions? Current governance frameworks assume a human actor whose rights and obligations are legally defined. They have no equivalent framework for AI agents. Without a formally complete authorization layer encoding permissions, obligations, and prohibitions at the agent level, AI agents operating in legal, financial, medical, and governmental contexts will act outside any enforceable rule structure. This is not a safety question - it is a question of legal personhood, authorization, and accountability that existing governance categories do not reach. Both issues require formal rule infrastructure as their technical foundation.

In the legal technology sector and across Europe, the governance gaps in our four priority areas are creating both acute risks and a narrow window of opportunity. The trustworthy AI gap is already causing harm in our sector. European law firms and legal departments are deploying AI tools that fabricate legal authority. Unlike the United States, where over 600 documented court cases have exposed this failure publicly, European jurisdictions have less systematic tracking — meaning the problem is likely larger than recorded. The absence of a formal correctness standard means every AI legal tool operates without a verifiable floor of reliability. Clients, courts, and regulators have no basis for trust. The transparency and accountability gap is acute for regulated industries. Financial institutions, healthcare providers, and public administrations across Europe are subject to GDPR, the AI Act, and sector-specific regulation — yet the AI systems they deploy cannot demonstrate formal compliance with the rules they are supposed to enforce. Accountability frameworks exist on paper; the technical infrastructure to satisfy them does not. This creates legal exposure at institutional scale. The human rights gap is being partially addressed by European regulation but not yet by technology. The EU Digital Identity Wallet mandate — requiring personal and business wallets across all 27 member states by end of 2026 — is the most ambitious rights-enabling digital infrastructure initiative globally. It encodes the principle that individuals control their own data. But the rule engine that makes consent, data sharing, and rights exercise machine-executable does not yet exist at scale. The regulatory framework is ahead of the technical infrastructure needed to fulfil it. The opportunity is time-bound. Europe is legislating faster than it is building. The 2026 wallet deadline, the AI Act enforcement timeline, and the acceleration of agentic AI deployment are converging simultaneously. Formal rule infrastructure developed now becomes the technical foundation for every governance framework that follows. The window to establish that foundation — before fragmented national implementations create incompatible standards — is open for approximately eighteen months. Nordic countries, with strong rule-of-law traditions, high digital literacy, and proximity to EU regulatory development, are uniquely positioned to lead on formal rule infrastructure for AI governance — not as recipients of frameworks designed elsewhere, but as technical architects of the standards the world will need.

The AI Dialogue occupies a position that no other institution currently holds: it is the only forum with the legitimacy, universality, and convening authority to establish shared foundations for AI governance that transcend regional regulatory competition. Its most important role is not producing more principles — it is translating existing principles into technical requirements. The world has no shortage of AI governance declarations. The Bletchley Declaration, the Hiroshima AI Process, the EU AI Act, the UNESCO Recommendation on AI Ethics, and dozens of national strategies share substantial common ground on values: safety, accountability, human rights, transparency. What is missing is the technical layer that makes those values enforceable. The Dialogue can mandate the work of converting governance principles into formal, machine-readable rule standards — the step that every existing framework stops short of taking. It can establish interoperability as a non-negotiable baseline. National AI governance frameworks are proliferating faster than they are converging. Without a shared formal rule language, AI systems will be governed differently across jurisdictions in ways that create regulatory arbitrage, undermine cross-border rights, and make accountability structurally impossible for systems that operate globally. The Dialogue can establish that any AI system operating across borders must satisfy a common formal rule layer — not identical regulation, but interoperable rule infrastructure. It can give voice to the jurisdictions most at risk from ungoverned AI. The countries with the least technical capacity are the most exposed to AI systems built elsewhere, governed by standards set elsewhere, and optimised for contexts other than their own. The Dialogue can ensure that formal AI governance standards are built with universal applicability from the outset — not retrofitted for the Global South after the architecture is fixed. The Dialogue succeeds if it moves the international community from shared values to shared infrastructure. That transition is where international cooperation becomes technically real.

Several existing initiatives provide foundations the Dialogue should explicitly connect with rather than duplicate. The EU AI Act and eIDAS 2.0 framework represent the most advanced binding regulatory infrastructure for AI governance globally. The AI Act establishes risk categories and conformity requirements; eIDAS 2.0 mandates personal and business digital identity wallets across all 27 member states by end of 2026. Together they create legal obligations without yet specifying the formal technical standards needed to satisfy them. The Dialogue should treat these frameworks as the floor, not the ceiling, and connect them to the global standard-setting bodies that can make them interoperable beyond European borders. The OECD AI Principles and the Global Partnership on AI (GPAI) have established shared normative language around trustworthy, human-centred AI across 46 countries. The Dialogue should build on this consensus rather than restart definitional debates, and focus GPAI's work specifically on the gap between principles and technically enforceable rule standards. ISO/IEC JTC 1/SC 42 — the international standards committee for AI — is developing technical standards for AI trustworthiness, bias, and transparency. The Dialogue should create a formal channel between political governance processes and this technical standardisation work, which currently proceeds largely independently of UN processes. The W3C and semantic web standards community has developed formal vocabularies for encoding rights, obligations, and permissions in machine-readable form — including ODRL, used in data governance. This technical infrastructure for rule encoding is mature and underutilised in AI governance discussions. The added value the Dialogue uniquely brings is convening authority across geopolitical divides that no regional or technical body possesses. The EU legislates; the OECD convenes democracies; ISO standardises technically. None can compel global interoperability. The Dialogue can establish the political mandate for a shared formal rule layer that makes AI governance frameworks mutually legible across jurisdictions — converting parallel national frameworks into a genuinely interoperable global system.

The Dialogue will succeed or fail based on whether it produces technically actionable outcomes rather than additional declarations. That requires a structure that integrates four distinct stakeholder contributions in sequence, not in parallel. Governments and intergovernmental bodies should open the Dialogue by mapping binding commitments already made — the EU AI Act, eIDAS 2.0, the Bletchley Declaration, national AI strategies — and identifying precisely where technical standards are absent. This frames the Dialogue around specific gaps rather than general principles, and prevents duplication of work already done. Technical and research communities — including formal methods researchers, standards bodies such as ISO/IEC JTC 1/SC 42 and W3C, and AI safety researchers — should present the state of existing technical infrastructure for rule encoding, formal verification, and agent authorization. Governance discussions routinely proceed without knowledge of what is already technically available. This step closes that gap and grounds political commitments in technical feasibility. Civil society and rights organisations should not be positioned as commentators at the margins but as co-authors of the authorization frameworks that determine what AI agents may do on behalf of individuals. The question of what rights people hold in relation to AI systems acting in their name — in legal, medical, financial, and governmental contexts — requires rights expertise at the table from the outset, not in consultation after frameworks are drafted. Private sector participants, including AI developers, legal technology companies, and infrastructure providers, should present concrete pilots and working implementations — not roadmaps. The Dialogue needs evidence of what formal rule infrastructure looks like when operational, not further proof-of-concept proposals. Structurally, the Dialogue should avoid plenary declarations as its primary output. Instead it should produce a small number of working groups with specific mandates, defined deliverables, and eighteen-month timelines aligned to the 2026 EU wallet deadline and AI Act enforcement — the nearest moment when governance infrastructure must be technically operational. Dialogue without deadlines produces documents. Deadlines produce infrastructure.

Four communities are systematically underrepresented, and their absence is not incidental — it shapes what governance frameworks are capable of addressing. Legal professionals and the judiciary. Judges, court administrators, and practising lawyers are experiencing the consequences of ungoverned AI deployment in real time — fabricated citations entering proceedings, AI-drafted documents misleading courts, algorithmic systems influencing legal outcomes. Yet global AI governance discussions are dominated by technologists, policymakers, and ethicists. The people closest to the rule-of-law failures AI is already causing have no structured voice in the governance processes meant to prevent them. Bar associations, judicial councils, and international legal bodies should be formal participants, not occasional consultants. Formal logicians, mathematicians, and legal theorists. The technical communities that understand what it means to reason correctly over rules — as distinct from predicting plausible text — are almost entirely absent from governance discussions. AI governance is being designed by people who understand AI as it is currently built, not by people who understand what AI would need to be built differently. This produces governance frameworks that regulate outputs without addressing the structural causes of failure. Bridging computer science, formal logic, and legal theory requires deliberately creating forums where these communities work together. Civil law and non-common-law jurisdictions. Global AI governance discussions disproportionately reflect common law assumptions — adversarial proceedings, case-law reasoning, judicial discretion. Civil law systems, Islamic legal traditions, customary law frameworks, and hybrid jurisdictions represent the legal reality of most of the world's population. AI systems trained and governed within common law assumptions will fail systematically when deployed elsewhere. Legal pluralism must be a design input, not an afterthought. Individuals whose rights are already being adjudicated by AI. Benefits recipients, asylum seekers, criminal defendants, and others subject to automated decisions affecting fundamental rights are the people governance frameworks are meant to protect. Participatory mechanisms — structured consultations, lived-experience advisory panels, accessible submission processes — must be built into governance processes by design, not added symbolically after frameworks are agreed. Inclusion is not representation in the room. It is representation in the architecture of what gets decided.

Three engagement formats would make the Dialogue substantively different from previous high-level AI convenings. Live adversarial testing of governance claims. The Dialogue should include structured sessions where AI systems are demonstrated — in real time — attempting tasks that current governance frameworks claim to address: legal reasoning, rights verification, regulatory compliance checking. Participants would observe directly where systems succeed and where they fail, grounding abstract governance debate in observable technical reality. Governance discussions that never touch a running system produce frameworks that fit no system. Policymakers, technical experts, civil society, and affected communities in the same room watching the same demonstration creates shared epistemic ground that position papers cannot. Cross-jurisdictional rule translation workshops. Rather than presenting national AI governance approaches in parallel, the Dialogue should run working sessions where participants attempt to translate a specific governance requirement — say, the right to explanation under GDPR, or an AI Act conformity obligation — into a machine-executable rule that a system from a different jurisdiction could also satisfy. The exercise exposes, concretely, where frameworks are genuinely compatible and where they diverge. It shifts the conversation from political alignment to technical interoperability, which is where the real governance gap lives. A structured dialogue between lawmakers and technical architects — with civil society as adjudicator. Most AI governance forums separate the political from the technical. The Dialogue should explicitly pair the authors of governance frameworks with the engineers building the systems those frameworks govern, with affected communities — particularly from the Global South, where AI governance capacity is most constrained — assessing whether the resulting commitments are meaningful. Power asymmetries in AI governance are not only between states but between those who write rules and those who build systems. This format makes that asymmetry visible and creates accountability within the Dialogue itself. The Dialogue succeeds if participants leave having changed their view of what is technically possible, not merely having reaffirmed what is politically desirable.

Five examples offer concrete models for effective AI governance, each addressing a different dimension of the challenge. The EU AI Act's risk-based conformity framework is the most structurally rigorous binding governance instrument deployed at scale. Its tiered approach - prohibiting certain uses, requiring conformity assessment for high-risk systems, and mandating transparency for limited-risk systems - establishes that AI governance can be legally enforceable, not merely aspirational. Its limitation is that conformity is assessed against documentation, not against formally verified system behaviour. The next evolution is conformity assessment against machine-executable rule standards. The EU Digital Identity Wallet mandate (Regulation EU 2024/1183) demonstrates that rights can be encoded as technical infrastructure. By requiring personal and business wallets across 27 member states by end of 2026, the EU has committed to making data rights - consent, access, portability, erasure - digitally exercisable rather than merely legally guaranteed. This is governance that builds the tool alongside the right. The OECD AI Principles implementation framework shows how normative consensus can be translated into national policy across diverse jurisdictions without requiring legal harmonisation. Its monitoring mechanisms provide comparative data on implementation gaps that purely political processes cannot generate. KROG Rules - a formally complete rule system developed over 30 years - demonstrates that any law, contract, or regulation can be encoded in formal logic that AI systems reason over correctly. It provides the technical foundation that current governance frameworks assume but do not specify: a rule layer that is machine-executable, formally verifiable, and isomorphic across domains including law, contracts, and regulatory compliance. It has been validated against enterprise legal workflows and compiles to standard formats including Solidity, OWL/RDF, and ODRL. Singapore's Model AI Governance Framework offers a practical, sector-specific implementation guide that bridges the gap between high-level principles and operational practice - demonstrating that governance can be both rigorous and implementable by organisations without deep technical capacity. Effective governance combines binding obligation, technical infrastructure, normative consensus, formal rule standards, and practical implementation guidance. No single initiative provides all five. The Dialogue's contribution is connecting them.