Centre for Digitial Ethics

Four outcomes would represent genuine success. 1. Formal recognition that international human rights law is the non-negotiable normative foundation for all subsequent AI governance work, and not one thematic area among many. This requires integrating the OHCHR, Special Procedures, and treaty body mechanisms into the Dialogue's ongoing architecture. The risk is not that human rights will be rejected, but that they will be rhetorically affirmed while structurally marginalised: included within a thematic cluster rather than treated as the organising commitment of the entire enterprise. 2. An explicit commitment to a governance floor rather than a governance ceiling. Frameworks designed around the most technologically advanced jurisdictions and exported downward structurally exclude most of the world from meaningful compliance. A baseline of minimum protections which are accessible, implementable, and rights-grounded must take priority over another set of aspirational high-level principles. The CDE notes that the Co-Chairs' proposed thematic cluster on 'bridging AI divides' is well-conceived in this respect, but a governance floor is a stronger commitment than capacity-building: it establishes what protection is owed, not merely what assistance might be provided. We would like to see the outcome of the Dialogue ultimately culminate in the drafting of an international treaty on digital rights. 3. The inclusion of structural and relational harms rather than just discrete, individual, identifiable injuries within the Dialogue's working definition of AI risk. Harms to democratic discourse, epistemic autonomy, collective self-governance, and meaningful human agency are precisely the harms least visible to existing product-safety risk frameworks and most damaging at scale. The proposed cluster on human rights and oversight provides a vehicle for this, but only if the Dialogue explicitly names structural and relational harms rather than defaulting to the individualised, event-based harm model that characterises existing frameworks. 4. A clear mandate to scope a cross-border AI incident monitoring and response mechanism including for agentic AI, as a concrete deliverable for the 2027 New York Dialogue. There is currently no global infrastructure for identifying, tracking, or responding to transboundary AI harms. The Draft Note's structure for Day 2 on 'Reflections, Convergences and the Way Forward', provides the ideal procedural vehicle for mandating this as a named, measurable commitment with accountable follow-through. Process legitimacy will depend on continuity and measurability. The co-chairs' summary should therefore go beyond a diplomatic record to identify named gaps, accountability holders, and a transparent roadmap to the 2027 New York Dialogue that civil society and affected communities can independently track and engage with in concrete ways. The CDE strongly endorses the April 23 consultation's focus on continuity mechanisms: standing thematic working groups and intersessional consultations are essential to ensure that the first Dialogue establishes a genuine architecture rather than an annual event.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights

The CDE selects the following two priorities as reflecting our most urgent concerns: Protection and promotion of human rights; and Social, economic, ethical, cultural, linguistic and technical implications of AI. The Centre for Digital Ethics works at the intersection of liberal political philosophy, legal theory, and digital regulation. Our selection reflects a conviction that AI governance cannot be reduced to technical risk management or procedural safeguards. Instead, it must be oriented toward the protection of human autonomy, dignity, and democratic self-governance. Protection and promotion of human rights is our primary selection because international human rights law supplies the only universally legitimate normative foundation for AI governance. It is the framework through which structural and relational harms - harms to democratic discourse, epistemic autonomy, collective agency, and meaningful human participation can be recognised, and not merely categorised as individual or transactional injuries. Without this foundation, governance frameworks risk becoming instruments of the most powerful states and corporations rather than protections for those most structurally exposed to AI-related harm. Social, economic, ethical, cultural, linguistic and technical implications of AI is the necessary companion priority. AI is not a neutral technical artefact. Rather it is reshaping labour markets, public discourse, cultural production, linguistic representation, and the material conditions of human flourishing. Governance that treats these dimensions as downstream consequences rather than as constitutive concerns will systematically underprotect the very populations that the Dialogue's inclusion commitments rhetorically centre. Taken together, these priorities express a single conviction: that the Dialogue's legitimacy depends on its willingness to engage with AI as a profoundly normative, political, and structural phenomenon rather than as a bounded technical risk. A governance architecture grounded in human rights and attentive to AI's full social, cultural, and economic footprint is one in which autonomy, democratic agency, and human flourishing can be meaningfully preserved. Anything narrower will reproduce, not remedy, existing asymmetries of power. The CDE observes that the Co-Chairs' proposed clustering merges human rights with transparency and oversight. While the CDE understands the structural logic, we urge the Co-Chairs to ensure that human rights is not reduced to a procedural obligation in this framing.

The CDE draws the Dialogue's attention to the following, each of which the preparatory process has either underweighted or left entirely unaddressed. Structural and relational harms: existing AI risk frameworks remain oriented toward individual, identifiable injuries such as data breaches, discriminatory outputs, product failures. They are poorly equipped to recognise harms that operate at the level of relationships, institutions, and collective self-governance: the erosion of epistemic autonomy, the degradation of democratic discourse, the manipulation of attention and preference formation, and the asymmetric reshaping of labour, cultural, and linguistic production. These are harms to the conditions of human flourishing itself and will be systematically under-regulated unless the Dialogue explicitly names them. They do not fit neatly within any of the four proposed clusters; they require explicit cross-cutting recognition. Concentration of power: governance frameworks that focus on AI systems without addressing the compute, data, energy, and capital stacks beneath them address symptoms rather than causes. The extraordinary concentration of foundation model development, compute capacity, and cloud infrastructure in a handful of firms and jurisdictions is itself a cross-cutting governance problem. Cross-border liability and access to remedy, anchored in the UN Guiding Principles on Business and Human Rights, require corresponding attention alongside state-level obligations. Agentic AI and autonomous systems: The rapid emergence of AI systems that act, transact, and execute on behalf of users, at scales no human principal can meaningfully oversee, raises distinct questions of attribution, liability, and human oversight that existing frameworks do not answer. This is not a horizon risk. Agentic AI systems are already operating at the operating system level in commercial deployment. The Dialogue must place this explicitly on the Geneva agenda rather than treating it as a future concern. Frontier capability and advanced AI risk: capability advances approaching or exceeding human-level performance in strategic domains raise distinct questions of alignment, oversight, and catastrophic risk that existing frameworks do not answer. The Dialogue should scope red lines, evaluation obligations, and deployment governance for the most capable systems, building on the International AI Safety Report process and the network of AI Safety Institutes. Anthropic's recent decision not to release Claude Mythos Preview on the basis that the model can autonomously identify unknown vulnerabilities and generate working exploits illustrates the scale and immediacy of this governance gap. Critical infrastructure and cyber-resilience: AI-enabled attacks on financial systems, energy grids, health systems, and public administration are materialising now. Governance must address transboundary cyber-resilience and deployment thresholds for frontier systems with offensive cyber capabilities. This requires formal coordination with the existing cybersecurity and international security architecture. Military and dual-use AI: AI in armed conflict, targeting, and autonomous weapons systems sits largely outside the Dialogue's current framing, despite profound implications for international humanitarian law, civilian protection, and human dignity. The Dialogue should engage explicitly with the CCW Group of Governmental Experts process on Lethal Autonomous Weapons Systems (LAWS) rather than treating military AI as out of scope. Environmental and intergenerational justice: The compute, water, and energy footprint of AI, and its implications for climate commitments and future generations, is inadequately reflected in the current thematic framing. Corporate accountability and access to remedy: state-centric governance leaves a significant gap where the most consequential actors are in fact private.

Ireland and the European Union: Ireland occupies an anomalous position as the European headquarters for most major US technology companies. The Irish Data Protection Commission and the emerging AI regulators function as de facto frontline enforcers for hundreds of millions of EU citizens. The structural gap is consistent: regulatory ambition at EU level outpaces enforcement capacity at member state level. The EU AI Act, the Digital Services Act, and the Digital Markets Act collectively represent the most developed rights-grounded framework globally, yet their legitimacy depends on avoiding the enforcement deficit that characterised GDPR's first decade. Ireland's experience as a host jurisdiction for AI infrastructure with all the sovereignty, liability, and environmental challenges that it entails offers the Dialogue direct, practical insight into the governance gaps that no framework currently addresses. The Transatlantic Divergence: from the CDE's Washington D.C. base, the asymmetry is acute. The current US policy trajectory has moved sharply toward voluntary, industry-led governance, while foundation model development, compute infrastructure, and agentic AI deployment remain overwhelmingly concentrated in US firms. The jurisdictions most exposed to AI-related harm have the least regulatory leverage over the actors producing it, and transboundary remedy mechanisms are correspondingly weak. This asymmetry is the central governance challenge that the Dialogue must name directly. The Sector: for civil society organisations working at the intersection of human rights, legal philosophy, and digital regulation, the most significant opportunity lies in the Global Dialogue's potential to establish a universally legitimate rights baseline. The most significant challenge is ensuring that the multilateral process does not become a forum for regulatory arbitrage where the least protective standards become the global norm through the practical leverage of the most powerful actors. A governance floor anchored in international human rights law and attentive to structural and relational harms is the only credible safeguard against that outcome.

The Dialogue's distinctive value lies precisely in what no other forum currently offers - a universal, UN-anchored, multistakeholder space in which AI governance can be addressed as an integrated normative question rather than through the fragmented lenses of trade, security, intellectual property, or sectoral regulation. Its role should be understood in those terms. It should not seek to replicate the technical evaluation work of the AI Safety Institute network, the standard-setting work of the Council of Europe or UNESCO, or the economic governance work of the OECD and G7. It should be the architecture that connects those efforts, confers universal legitimacy on shared commitments, and ensures that those most exposed to AI-related harm are co-producers of governance rather than consultees. The CDE identifies five distinctive cooperative roles. Establishing a common normative floor: the Dialogue is uniquely positioned to consolidate international human rights law as the shared baseline for AI governance. No regional bloc, industry forum, or bilateral arrangement can confer universal legitimacy. Only the UN can. The Dialogue must use that legitimacy to establish commitments that are binding in character, not merely aspirational. Providing coherence without duplication: a proliferation of overlapping initiatives risks regulatory fragmentation and forum-shopping. The Dialogue should serve as the coordinating architecture that maps, harmonises, and translates these efforts into interoperable standards, rather than producing yet another parallel track. Giving structural voice to those most exposed and least represented: the Dialogue can institutionalise meaningful participation for Global South states, affected communities, and civil society not as consultees but as co-producers of governance. Without this, international cooperation reproduces existing asymmetries of power. This is the purpose of the governance floor: to ensure that the rules that govern AI are set by those who will live under them, not only by those who produce and deploy the technology. Scoping concrete cross-border deliverables: the Dialogue should catalyse mechanisms that no single jurisdiction can build alone: a shared AI incident monitoring and response infrastructure; interoperable evaluation and red-teaming standards for frontier systems; coordinated deployment thresholds for systems with offensive cyber or military applications; and cross-border remedy pathways anchored in the UN Guiding Principles on Business and Human Rights. These should be named as mandated deliverables for the 2027 New York session in the Co-Chairs' summary. Sustaining continuity: the Dialogue's most enduring contribution may be procedural: a durable, predictable, and measurable cooperation architecture that outlasts political cycles and corporate product cycles alike. The April 23 guiding question on continuity mechanisms is therefore one of the most important questions the Dialogue faces. Standing thematic working groups, intersessional consultations, and a permanent multistakeholder advisory forum are the minimum requirements for a governance architecture capable of matching the pace of the technology.

The Dialogue should anchor itself in the UN human rights system: the OHCHR, the Human Rights Council, relevant Special Procedures and most notably the Special Rapporteurs on freedom of opinion and expression, privacy, cultural rights, and the rights of persons with disabilities and the treaty body mechanisms. These provide the normative foundation that no other forum can replicate. A UN Special Rapporteur on AI and Human Rights. The CDE, in alignment with the Center for AI and Digital Policy (CAIDP), strongly endorses the creation of a dedicated Special Rapporteur mandate under the Human Rights Council. An independent expert with the agility, authority, and focus to monitor AI-related human rights impacts, engage with states through the Universal Periodic Review, and drive incident-reporting mechanisms would give the Dialogue a permanent, authoritative counterpart between sessions. This is the single most consequential institutional reform available to the Dialogue and the CDE urges the Co-Chairs to include it in their summary. Standard-setting instruments: UNESCO's Recommendation on the Ethics of AI, the Council of Europe Framework Convention on AI, Human Rights, Democracy and the Rule of Law, the OECD AI Principles, and the EU AI Act together represent the most developed rights-grounded framework. The Dialogue should treat these as baseline architecture to harmonise rather than duplicate. Safety and evaluation infrastructure: the International AI Safety Report process and the network of AI Safety and AI Security Institutes (UK AISI, US NIST AISI, and counterparts) offer technical evaluation capacity the UN lacks. Formal interoperability between these bodies and the Dialogue is essential, particularly for frontier model evaluation and red-teaming standards. Complementary UN and multilateral processes: The Global Digital Compact, the IGF, WSIS, ITU, the CCW GGE on LAWS, and the Independent International Scientific Panel on AI all intersect with the Dialogue's mandate. Clear complementarity, not overlap, is critical. The Dialogue's Day 2 'Dialogue of Dialogues' session is well-positioned to map these relationships explicitly. Business and human rights: The UN Guiding Principles on Business and Human Rights and the Working Group on Business and Human Rights supply accountability infrastructure for private-sector actors that state-centric frameworks lack. The Dialogue should formally integrate this framework rather than treating corporate accountability as a secondary concern. Civil society evidence and monitoring: CAIDP's AI and Democratic Values Index, Ranking Digital Rights, the AI Incident Database, and the Global Call for AI Red Lines represent hard-won evidence and advocacy infrastructure. These should be formally integrated into the Dialogue's evidence base rather than treated as peripheral inputs.

1) Adopt a continuous model rather than an annual summit with sporadic engagement. Standing thematic working groups, intersessional consultations, and a permanent multistakeholder advisory forum are the minimum requirements for a governance architecture capable of matching the pace of the technology. 2) Ensure structured parity. Civil society and Global South stakeholders must have guaranteed deliberative roles, not invited slots. The Co-Chairs' proposed structure of co-chairing thematic discussions with a Member State and a relevant stakeholder is a welcome step; it should be extended to the plenary segments. 3) Commit to procedural transparency: open agendas, published inputs, traceable influence, and a participant register disclosing material interests. A public-facing portal showing which submissions and interventions shaped which outcomes would close the accountability loop and incentivise substantive participation. 4) Provide meaningful resourcing including travel, translation, technical support where relevant and necessary so that participation does not depend on institutional wealth. 5) Establish regional consultation hubs to surface diverse inputs, with particular attention to Africa, Latin America, the Caribbean, Asia-Pacific, and Small Island Developing States feeding directly into plenary sessions, with real-time hybrid participation and asynchronous input channels for those unable to travel. 6) Build in accountability loops so each Dialogue cycle can report publicly on which inputs shaped outcomes, and which were set aside and why. Inclusion should be measurable, traceable, and consequential or it risk looking performative.

AI governance discussions are overwhelmingly shaped by governments of the Global North, a small number of multinational technology firms, and the international civil society organisations best resourced to engage with them. Several voices are systematically absent:- Global South states and communities: countries where AI systems are deployed, tested, and scaled but rarely designed or governed have the least formal voice in shaping the rules applied to them. Inclusion requires more than invitation; it requires funded participation, regional consultation hubs, and weighted deliberative structures. Indigenous peoples: questions of data sovereignty, cultural representation in training data, linguistic preservation, and consent to extraction and use of traditional knowledge are largely absent from mainstream AI governance. Indigenous representative bodies should have guaranteed participation, drawing on the precedent set by the UN Permanent Forum on Indigenous Issues. Workers and trade unions: AI's reshaping of labour markets from algorithmic management to job displacement is a defining social consequence, yet workers' voices are marginal. The ILO should be formally integrated into the Dialogue's architecture, and trade union confederations should be given standing in thematic working groups. Women, girls, and gender-diverse communities: technology-facilitated gender-based violence, biased training data, and the gendered distribution of both AI harms and benefits warrant dedicated governance attention. UN Women and the Commission on the Status of Women should be embedded structurally in the Dialogue's architecture. Children and young people: those who will live longest under AI-governed conditions have the least voice in shaping them. UNICEF-supported child participation mechanisms should be formalised, and the Ombudsman for Children and equivalent bodies in other jurisdictions given standing to submit evidence. Persons with disabilities: despite the CRPD framework, disability perspectives remain peripheral. The Committee on the Rights of Persons with Disabilities should be formally integrated into the Dialogue's human rights architecture. Linguistic and cultural minorities: the dominance of English and a handful of major languages in AI training data and deployment structurally excludes most of humanity's linguistic diversity. Environmental and future-generations advocates: intergenerational justice requires formal representation, not only rhetorical acknowledgment. The Dialogue should establish a formal mechanism for intergenerational input.

Traditional plenary formats with prepared statements delivered sequentially to a passive audience tend to produce limited deliberative value. The CDE recommends that the first Dialogue experiment with formats that generate genuine exchange, surface disagreement productively, and translate inputs into outcomes. Deliberative mini-publics: citizens' assemblies can bring lay public voice into the Dialogue on specific questions for example, acceptable uses of AI in public services or agentic AI deployment thresholds. Structured scenario exercises: red-teaming workshops, horizon-scanning panels, and cross-sectoral tabletop exercises for example, simulating responses to a cross-border AI-enabled cyber incident, a frontier capability jump, or an algorithmic election interference event, would surface governance gaps more effectively than abstract principle debates. Evidence-led working groups: standing thematic groups with embedded technical, legal, and lived-experience expertise should produce published working drafts between plenary sessions, enabling iterative refinement rather than episodic engagement. Moderated cross-examination panels: Industry, civil society, Global South states, and affected communities should be placed in structured exchanges with genuine follow-up questioning, rather than sequential monologues. Moderators should be drawn from independent academic or judicial backgrounds. Youth and intergenerational panels: formal intergenerational dialogues giving young people and future-generation advocates deliberative standing, not ceremonial addresses.

Binding rights-grounded instruments: the Council of Europe Framework Convention on AI, Human Rights, Democracy and the Rule of Law (2024), the EU AI Act, the Digital Services Act, the Digital Markets Act, and the revised EU Product Liability Directive (2024/2853) (which extends liability to software and AI systems, lowers evidentiary barriers for claimants, and recognises psychological harm as a compensable injury) together constitute the most developed rights-based regulatory architecture and offer a universal template. The inclusion of psychological harm within the Product Liability Directive is particularly significant: to our knowledge it represents the first binding instrument to formally recognise the category of non-material AI-mediated harm. Official harms frameworks: in its submission to the Irish Government on the Regulation of Artificial Intelligence Bill 2026, the CDE has proposed a statutory duty on the national AI regulator to develop and periodically update an official guidance framework on AI-related harms which explicitly encompasses structural, psychological, epistemic, relational, and democratic harms alongside physical and material harms. Risk frameworks derived from product safety law are structurally inadequate for ambient, cumulative, and collective harms. An official harms typology offers a scalable corrective that the Dialogue should adopt as a model for international guidance. Fundamental Rights and Democracy Impact Assessments: the Dutch FRAIA model, the EU AI Act's Fundamental Rights Impact Assessment obligation, and the CDE's proposed alignment with Ireland's Public Sector Human Rights and Equality Duty (s.42 IHREC Act 2014) operationalise rights protection procedurally before deployment rather than through post-hoc enforcement. A parallel Democracy Impact Assessment framework for AI systems operating in the political information environment is a necessary extension and one with no current international equivalent. Enhanced scrutiny for rights-sensitive AI: extending enhanced scrutiny beyond product-safety categories to any AI system materially affecting rights, livelihoods, or access to essential services including employment, health, dynamic pricing, welfare, and procurement, delivers a genuine governance floor. The MiFID II suitability assessment framework in EU financial services provides a tested analogy for AI deployment accountability: requiring deployers to assess whether AI systems are appropriate for the specific populations they affect. Child-specific safeguards: child Rights Impact Assessments aligned with the UNCRC and UNICEF guidance; safety-by-design obligations; age-verification requirements for AI chatbots and companion systems; and a dedicated regulatory category for AI companion systems and AI toys represent the minimum architecture for child protection. The CDE's submission to the Irish AI Bill elaborates this framework in detail; the Dialogue should adopt equivalent international standards. The Safety, Evaluation, and Enforcement Infrastructure must shift from voluntary pacts to a statutory global architecture to include AI regulatory bodies. By leveraging the technical capacity of safety institutes, the Dialogue could seek the establishment of a licensing regime requiring pre-deployment attestation by engineers and universal red-teaming for frontier models. This technical layer would be backed by the enforcement power of bodies like the EU AI Office capable of mandating audits and market withdrawals. The framework could also incorporate individual accountability for negligent/deliberately unsafe deployment, potentially resulting in the loss of employment. Deliberative democracy: the Citizens' Assembly model provides a proven mechanism for structured public deliberation on complex governance questions. Civil society monitoring and accountability: these should be formally integrated as reference infrastructure rather than treated as external submissions.