University of Carthage-Tunisia

The first Dialogue will be a success if it establishes what multilateral AI governance currently lacks: a shared normative reference framework that distinguishes between what is already legally binding and what remains to be negotiated. The most concrete contribution the Dialogue could make is the explicit mapping of existing law applicable to AI systems: identification of obligations under general international law (jus cogens, State responsibility, due diligence) already binding upon all 193 Member States, applicable to algorithmic harm regardless of any new treaty. This work transforms the Dialogue from a platform for exchanging principles into a deliberation grounded in verifiable, pre-existing obligations. It also provides the only durable basis for measuring whether Global South priorities are reflected in outcomes, not by the composition of the room, but by their presence in the co-chairs' summary. Beyond the scheduled 2026–2027 sequence, what the Dialogue needs to produce is a substantive inter-sessional work mandate: thematic working streams with defined guiding questions, and structured contribution mechanisms for the Independent International Scientific Panel that go beyond its annual report presentation. A calendar is not a roadmap. What the Dialogue must avoid is equally clear: a co-chairs' summary so consensual it fails to resolve anything. The value of a non-binding forum is precisely the freedom to state difficult truths without the constraint of a vote i.e. a freedom whose value depends entirely on whether it is exercised.

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

The research programme developed over two decades in public international law, AI governance, and human rights maps directly onto four of the five thematic areas identified in resolution A/RES/79/325, and does so not thematically, but structurally: each area corresponds to a distinct layer of the same legal argument. The foundational layer is the protection and promotion of human rights. Non-derogable norms (right to life, prohibition of systemic discrimination) define the legal threshold below which no AI governance framework can operate. This is a matter of existing obligation, not policy preference, and it is from this threshold that the remaining questions derive their legal force. Safe, secure and trustworthy AI is where that threshold meets State conduct. The central question of our programme "whether States' due diligence obligations extend to high-risk AI systems deployed under their jurisdiction" follows directly from this human rights foundation. The doctrine of responsible State behaviour in cyberspace, developed through the UN GGE and OEWG processes, provides a transposable doctrinal framework, making systemic security, in this reading, an obligation of State conduct. Transparency and accountability is where the obligation meets its limits. Algorithmic opacity makes accountability structurally elusive in a way that without opposable traceability obligations, attribution of transboundary harm remains theoretical regardless of the norms invoked. The ILC Articles on State Responsibility (2001), and their extension to private digital actors operating under State jurisdiction, are the relevant doctrinal axis, precisely because they address the gap between the existence of an obligation and the capacity to enforce it. Interoperability of governance approaches is where the argument reaches its practical conclusion. Jus cogens constitutes the only reference framework that all regulatory regimes (the EU AI Act, national MENA legislation, bilateral digital agreements) are already legally bound to respect. Interoperability anchored at the top of the normative hierarchy is more durable than convergence negotiated from the bottom, because it does not depend on political agreement but it already exists.

Two substantive issues remain outside the listed thematic areas. Cognitive sovereignty is the first gap since generative AI systems are altering the conditions under which opinions form, information circulates, and autonomous reasoning is exercised in digital public spaces. This is not a question of transparency or ethics, it goes to the structural preconditions of informed consent and democratic deliberation. Freedom of thought, opinion, and expression are first-generation human rights obligations. They are materially affected when informational infrastructure is privately owned and algorithmically curated. The issue does not require new norms; it requires applying existing ones to a configuration they were not designed to anticipate. The privatisation of State functions through AI outsourcing is the more operationally urgent gap. When States delegate essential public services ( judicial decisions, border management, social benefit allocation) to private AI providers, the chain of legal accountability fractures in ways that existing doctrine has not resolved. The ILC Articles on State Responsibility were not designed for a configuration in which a State acts through a private algorithmic intermediary it neither fully controls nor fully audits. Whether and how State responsibility attaches in these conditions is an unresolved doctrinal question that cuts across every thematic cluster the Dialogue has identified. It means the Dialogue will address AI governance while remaining silent on two of its most consequential structural effects.

Tunisia and the MENA region sit at the intersection of the governance gaps this Dialogue addresses, not as passive recipients, but as jurisdictions where the absence of an operational normative framework produces measurable legal consequences. The most significant challenge is asymmetric regulatory exposure. Tunisian institutions and citizens are de facto subject to the extraterritorial reach of the US CLOUD Act and the EU AI Act without having participated in their negotiation. Our existing legal infrastructure (including Convention 108+, the Budapest Convention, and national digital decrees) cannot unilaterally resolve cannot unilaterally resolve the jurisdictional conflicts that arise when transboundary algorithmic harm occurs. The accountability chain fractures precisely where systems developed and owned outside the jurisdiction produce legal effects within it. Simultaneously, we face the privatisation of public functions through AI outsourcing. Essential services ( border management, social safety nets, and judicial support) are increasingly mediated by private algorithmic intermediaries. This creates a systemic accountability gap that technical training alone cannot bridge; it is an unresolved doctrinal question regarding the delegation of sovereign authority to non-state technical actors. The opportunity is specific: Tunisia's legal infrastructure (Convention 108+, Budapest Convention adhesion, an active domestic legislative and regulatory process) positions it as a credible normative laboratory. We offer a specific empirical environment to test how general international law obligations can be operationalised within a middle-income country's regulatory framework. By bridging the gap between existing international obligations and their algorithmic application, Tunisia can provide a template for interoperability that aligns universal human rights obligations with meaningful regulatory autonomy over AI systems operating within national jurisdiction.

The Dialogue's most consequential contribution would be one that no existing forum is structurally positioned to make: establishing the normative relationship between existing binding international law and the emerging body of AI governance instruments. This is not a gap in political will, it is a gap in legal architecture, and it compounds with every new soft law instrument adopted in the absence of a common reference framework. Concretely, the Dialogue can advance international cooperation. It must function as an interpretive forum rather than a negotiating one, using its non-binding character not as a limitation but as a licence to address questions that formal treaty bodies cannot yet reach. It must produce outputs that are usable: not declarations of principle, but structured analyses of how existing obligations ( due diligence, State responsibility, non-derogable human rights norms) apply to specific configurations of algorithmic harm. And it must build the kind of institutional memory that allows the 2027 session to advance from where 2026 ends, rather than restarting the same foundational conversation. The role the Dialogue cannot play, and should not attempt, is that of a regulatory body. Its value lies precisely in the space between political consensus and legal obligation: mapping what already binds States, identifying where binding frameworks are absent, and building the shared understanding that makes future binding commitments politically possible. International cooperation on AI governance does not currently lack forums. It lacks a forum where the legal architecture is treated with the same seriousness as the technical one.

The landscape the Dialogue enters is not empty. The UNESCO Recommendation on the Ethics of AI (2021) provides the most universally adopted ethical framework, with implementation mechanisms (including the ROAM-X methodology and the Global AI Ethics and Governance Observatory) that offer a direct operational interface between normative commitments and national policy processes. The Council of Europe Framework Convention on AI (CETS No. 225) represents the first binding multilateral instrument, establishing a model for how treaty-based obligations can coexist with the flexibility required by fast-moving technology. The UN GGE and OEWG processes on responsible State behaviour in cyberspace have produced the most developed doctrine on due diligence obligations in digital environments, a doctrine that is directly transposable to AI systems and has been systematically underutilised in AI governance discussions. The ILC Articles on State Responsibility (2001) remain the authoritative framework for attribution of transboundary harm and have yet to be seriously applied to algorithmic contexts. What the Dialogue adds is not another instrument. It adds the institutional space under General Assembly authority, with universal membership, where these frameworks are brought into explicit relationship with one another. The EU AI Act, the CETS No. 225, the UNESCO Recommendation, and bilateral digital agreements currently operate in parallel without a common normative hierarchy. The Dialogue is the only forum where that hierarchy can be mapped without reopening any of the underlying instruments. Its specific added value for States with limited regulatory capacity is access: access to the interpretive work being done in Geneva, Brussels, and Strasbourg, translated into terms that national legislators, courts, and procurement bodies in the MENA region and the Global South can directly apply. The Dialogue does not need to produce new law. It needs to make existing law legible to those who need it most.

The Dialogue's multistakeholder design is its most valuable institutional feature and its most significant operational risk. Without a structured contribution architecture, multistakeholder participation defaults to a format where the best-resourced actors dominate and procedural inclusion substitutes for substantive influence. Written submissions should be treated as primary input, not background documentation. The current consultation process generates substantive legal and technical analysis that disappears into synthesis documents. A more productive model would have thematic rapporteurs, drawn from academia and civil society rather than solely from secretariat staff, responsible for presenting submission clusters directly in plenary, with right of response from submitting parties. This gives the submission process operational weight rather than procedural form. Academic and research institutions should be engaged not as observers but as the Dialogue's analytical infrastructure. The Independent International Scientific Panel cannot cover the full doctrinal terrain the Dialogue needs mapped. Commissioned working papers on specific questions such as the application of ILC Articles to algorithmic harm, or the extraterritorial reach of existing AI regulations, would give the Dialogue the technical depth that neither government delegations nor civil society organisations are positioned to provide alone. States with limited AI capacity require dedicated pre-session technical briefings as a structural condition of informed participation. A delegation that receives the agenda forty-eight hours before the session cannot contribute substantively to questions it has had no time to analyse. Regional preparatory consultations (the MENA process, the African Group process) should feed directly into the plenary, with their conclusions shaping the core agenda rather than being reported as separate, secondary documents Format recommendation: sequential plenary sessions with defined thematic blocks, each opening with a rapporteur synthesis and closing with a structured exchange rather than open-floor statements.

The underrepresentation that most affects the Dialogue's analytical quality is epistemic rather than demographic. The voices currently absent are not simply those of the Global South in general, but specifically those of legal scholars, judges, and practitioners from non-Western jurisdictions who work within international law frameworks and whose interpretive traditions differ systematically from those dominant in Geneva and Brussels. Francophone and Arabic-speaking legal scholars represent a body of international law doctrine, developed through the Institut de droit international, the Académie de droit international de La Haye, and regional judicial bodies that is structurally absent from AI governance discussions conducted primarily in English. Their exclusion is not incidental but it means that the normative frameworks being developed reflect common law and Northern European civil law traditions, while other law traditions and francophone civil law approaches to State responsibility, public order, and fundamental rights remain outside the interpretive conversation. The Dialogue should commission working papers specifically from scholars in these traditions and ensure Arabic and French are working languages in thematic sessions, not only official languages for ceremonial purposes. Domestic courts and administrative tribunals from middle-income countries are the institutions that will actually apply AI governance frameworks to concrete cases. Their perspective on what is operationally feasible (what attribution standards can be applied, what transparency obligations can be enforced, what due diligence looks like in a jurisdiction without a dedicated AI regulator) remains systematically absent from the Dialogue's current design. Academics and civil society actors working on algorithmic harm at the community level in the Global South possess the empirical evidence base that the Dialogue's scientific panel needs to ground its assessments. Their inclusion requires advance funding for participation, since accreditation without travel funding produces nominal inclusion.

The Dialogue should avoid replicating the most common format in multilateral AI governance : sequential three-minute statements from government delegations, followed by a panel of experts, followed by a co-chairs' synthesis that reflects neither. Structured legal argumentation sessions, modelled loosely on moot court or oral pleading formats, would require delegations and stakeholders to engage directly with specific normative questions : for example, whether States' due diligence obligations extend to AI systems deployed by private actors under their jurisdiction, and to respond to counter-arguments in real time. This produces clarity on where genuine disagreement lies, rather than the convergence of declared positions that open-floor formats tend to produce. It also creates a public record of positions that can anchor inter-sessional work. Scenario-based deliberation replaces abstract thematic discussion with concrete cases: a specific algorithmic harm, a specific jurisdictional configuration, a specific gap between existing law and current practice. Delegations reason from the particular to the general rather than declaring principles that evade operational content. The Independent International Scientific Panel is well-positioned to develop the scenario library. Inter-sessional working groups with defined deliverables, not standing committees, but time-limited groups tasked with producing a specific analytical output by a specific date, would transform the period between Geneva 2026 and New York 2027 from a gap into a production phase. Their outputs would constitute the substantive agenda of the second session rather than leaving it to be constructed from scratch. Adding remote access to a format that does not function well in person does not produce inclusion, it merely scales the underlying dysfunction.

Four examples illustrate what operational AI governance looks like when existing legal frameworks are applied seriously rather than supplemented with new principles. The GDPR adequacy mechanism, most recently instantiated in the EU-US Data Privacy Framework, demonstrates that regulatory interoperability does not require harmonisation of legal traditions. It requires demonstrated equivalence in protection outcomes. This model is directly transposable to AI governance: the Dialogue could develop equivalence criteria based on existing human rights obligations rather than negotiating common technical standards that no jurisdiction will uniformly adopt. Spain's AI regulatory sandbox, designed to pilot the EU AI Act before its full entry into force, converts transparency from a declared principle into a technical protocol with enforcement consequences. Requiring third-party algorithmic audits prior to market entry shifts the burden of proof from the victim of harm to the provider, a structural inversion that general product liability doctrine supports and that AI governance has been slow to operationalise. The UNESCO Readiness Assessment Methodology offers the most developed framework for grounding international AI governance commitments in national institutional realities. Unlike purely technical benchmarks, the RAM requires States to assess judicial, social, and linguistic capacity, producing the empirical baseline that makes international cooperation substantive rather than declaratory. Its application across MENA jurisdictions, including Tunisia, provides the Dialogue with comparative data that its scientific panel should be systematically using. For middle-income States, the more instructive governance models are those that treat domestic legal infrastructure as a regulatory asset rather than a deficit. Tunisia's combination of Convention 108+ ratification, Budapest Convention adhesion, RAM adoption, and active domestic digital legislation demonstrates that general international law instruments, when applied with interpretive seriousness, reach further into AI accountability than purpose-built AI regulation adopted without enforcement capacity.