University of Ljubljana

The first Global Dialogue will succeed if it produces outcomes that are concrete enough to be tracked at the second session in May 2027. Three criteria matter. First, measurable commitments rather than principles. The international AI governance landscape is not short on principles; it is short on uptake. The Dialogue should close with specific, attributable commitments from Member States, industry, and international organisations (training places funded, compute access opened, evaluation capacity hosted, datasets shared) reported against at the next session. Without this, the Dialogue risks duplicating the rhetorical layer already produced by the OECD, UNESCO, the Council of Europe, and the G7 Hiroshima Process. Second, a credible response to the capacity gap. Most non-frontier states cannot independently evaluate, audit, or red-team the systems they are asked to regulate. Until they can, safety, human-rights, and interoperability commitments rest on capacity that does not exist. A success criterion for July is the launch of an AI Capacity Compact, which is a multi-year, multi-stakeholder commitment to expand training, compute access, and evaluation infrastructure in under-resourced regions, with measurable targets. Third, interoperability work that is co-designed, not retrofitted. The Dialogue should produce a public interoperability map comparing the EU AI Act, the Council of Europe Framework Convention, the UNESCO Recommendation, and major national frameworks at the level of specific obligations. This makes gaps and conflicts visible and gives smaller jurisdictions an evidence base for their own choices, rather than presenting them with a fait accompli. Underpinning all three: the Independent Scientific Panel's report should be released at least four weeks before Geneva, not at it, so that delegations can engage with evidence rather than receive it. Success means the Dialogue is informed, operational, and trackable — not declarative.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• AI capacity-building
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

The four areas are connected by a single causal chain: capacity determines whether the other three are real. AI capacity-building is the binding constraint. Independent evaluation, red-teaming, audit, and incident analysis require people, compute, and tooling concentrated in a handful of jurisdictions. Most Member States cannot perform these functions on the systems they are asked to regulate. Without addressing this asymmetry, every other commitment degrades: oversight becomes formal rather than substantive, and interoperability becomes harmonisation-by-adoption. Social, economic, cultural, and linguistic implications are where the capacity gap is felt most acutely and most unevenly. Frontier models are trained predominantly on English-language and high-resource-language data, encoding the cultural assumptions of a small number of societies. For smaller linguistic communities, including my own (Slovenian), this is not a peripheral concern but a question of whether AI systems function adequately at all in the local context. The Dialogue should treat linguistic and cultural inclusion as a technical and infrastructural issue, not solely an ethical one. Interoperability of governance approaches matters because fragmentation is already advanced. The EU AI Act, the Council of Europe Framework Convention, the UNESCO Recommendation, and major national frameworks impose overlapping but non-identical obligations. Smaller jurisdictions face the cost of compliance with multiple regimes without the capacity to shape any. Interoperability must be co-designed, with a public map comparing instruments at the level of specific obligations, rather than retrofitted to states after the fact. Transparency, accountability, and human oversight are the operational test of whether the previous three are working. Oversight without technical access is performative. Accountability without cross-border incident reporting is territorial fiction. The Dialogue should advance shared minimum standards, that travel across jurisdictions.

Three cross-cutting issues are inadequately captured by the listed themes. First, the compute and energy substrate of AI governance. Discussions of safety, transparency, and capacity treat AI systems as if they exist in software alone, but they do not. Frontier model training and inference depend on physical infrastructure (semiconductors, data centres, electricity, water) concentrated in a small number of jurisdictions and increasingly subject to export controls and industrial policy. Any governance regime that ignores the compute layer governs the visible surface of the problem. The Dialogue should treat compute access, energy use, and supply-chain resilience as first-order governance questions, not technical footnotes. Second, the civilian-military boundary in AI development and deployment. Frontier AI laboratories increasingly hold contracts with defence and intelligence agencies. Dual-use deployment is not a future concern, but current practice. The listed themes treat AI governance as a civilian project, which understates the problem. The Dialogue need not resolve military AI questions, which are addressed in other forums, but it should at minimum establish that civilian governance pipelines and military procurement pipelines require structural separation, and that transparency commitments cannot be voided by national-security framing. Third, governance of AI agents and agentic systems. Existing instruments, including the EU AI Act, were drafted around models that produce outputs for human review. Agentic systems that take actions in digital and physical environments - booking, transacting, communicating, deploying code - fall awkwardly across current categories. However, liability, oversight, and incident reporting frameworks have not caught up. A fourth, briefly: child safety and AI, which the EU has flagged and which deserves explicit mention rather than absorption into general human-rights language.

Slovenia is a useful case because it sits at the intersection of three asymmetries the Dialogue must address. Capacity asymmetry within the EU: Slovenia is bound by the AI Act, GDPR, and Interoperable Europe Act, but its national capacity to operationalise them is structurally smaller than that of larger Member States. Notified bodies, conformity assessment infrastructure, and AI evaluation expertise are concentrated in a few capitals. For smaller states, compliance costs scale roughly with obligation, while capacity scales with population and budget. Without coordinated capacity-building, including the EU AI Office, AI Factories, and national hubs such as Slovenia's EDIH network, the AI Act risks producing uneven enforcement across the single market. Linguistic and cultural underrepresentation: Slovenian is a low-resource language by NLP standards. Frontier model performance in Slovenian lags substantially behind English, which directly affects public-service deployment, education, and accessibility. This is not solved by translation layers; it requires investment in language resources, evaluation benchmarks, and locally-trained or fine-tuned models. Without such investment, AI deployment in smaller linguistic communities reproduces the historical pattern of digital infrastructure being optimised for the largest markets. Sectoral exposure in the public sector and SMEs: As a former Minister of Digital Transformation, I have observed the same pattern across public administrations and SMEs i.e. appetite to adopt AI is high, but the capacity to evaluate vendor claims, audit deployed systems, and respond to incidents is not. The opportunity in terms of productivity gains, better public services, more competitive SMEs, etc. is significant, but it is contingent on independent evaluation capacity that does not currently exist at the necessary scale. The Dialogue should treat these patterns as representative, not exceptional. Most Member States face structurally similar conditions.

The Dialogue's role is to be the universal connective layer of an already crowded international AI governance landscape and not a new standard-setter. What it should not try to be: The Dialogue should not produce parallel principles, parallel risk taxonomies, or parallel evaluation frameworks. The OECD AI Principles, the UNESCO Recommendation, the Council of Europe Framework Convention, the G7 Hiroshima Process, ISO/IEC JTC 1/SC 42, and the EU AI Act already cover this ground. Adding a further substantive layer would deepen fragmentation, not reduce it, and would dilute the Dialogue's comparative advantage. What it should be: Three roles are distinctive and feasible: * The Dialogue is the only forum where all 193 Member States can deliberate together. Its function is therefore to surface the priorities of states currently outside the rule-making process, to bring those priorities into existing instruments rather than alongside them, and to make the trade-offs visible to all parties. * The Dialogue can function as the connective tissue between instruments. A maintained interoperability map, regular comparison of obligations across regimes, and a shared incident-reporting vocabulary aligned with the OECD AI Incidents Monitor, NIST AI RMF, and ISO/IEC 42001 would deliver concrete value without duplication. * Paired with the Independent Scientific Panel, the Dialogue can become the authoritative venue where governance discussion is anchored to current technical reality. Most international AI processes work from a state of the field that is twelve to twenty-four months out of date. A standing scientific input, pre-circulated and openly contested, would correct this. The test of success is simple: by the second session in May 2027, can a delegate from any Member State point to specific commitments, comparisons, and evidence the Dialogue produced and that they have used at home?

The existing landscape is best understood by function, not institution. The Dialogue should connect across five functional layers. * Standard-setting: OECD AI Principles, UNESCO Recommendation, Council of Europe Framework Convention, G7 Hiroshima Process, EU AI Act, ISO/IEC JTC 1/SC 42. These define obligations. * Technical evaluation: NIST AI RMF, MLCommons, the international AI Safety Institute network, Singapore's AI Verify, ISO/IEC 42001. These test whether obligations are met. *Capacity and adoption: GPAI, UNESCO Readiness Assessment Methodology, ITU AI for Good, the EU AI Office and AI Factories, the African Union Continental AI Strategy. These determine whether obligations can be met. * Incident monitoring: OECD AI Incidents Monitor, AI Vulnerability Database initiatives, voluntary frontier safety commitments. * Human rights: OHCHR B-Tech, UN Human Rights Council mechanisms, regional human rights bodies. The Dialogue's added value is to make these layers legible to one another: a maintained interoperability map across instruments, a shared incident-reporting vocabulary, a Capacity Compact connecting parallel programmes, and a structured channel for the Independent Scientific Panel to feed all five layers. This is connective infrastructure, not new substance, and it is what no single existing body has the universal mandate to provide.

Stakeholder contributions: Member States should bring measurable national commitments, including training places, compute access, evaluation capacity, and datasets opened, rather than principles statements, and report against them at the next session. Private sector should disclose, in standardised form, model evaluations, incident reports, and the geographic distribution of compute access programmes. Voluntary commitments should be reported against publicly. Civil society should prioritise affected-community evidence over abstract principles. The Dialogue already has a surplus of the latter. Academia and the technical community should provide independent evaluation methodology, standards bridges to IEEE, ISO/IEC, IETF and MLCommons, and training capacity. Universities outside frontier-AI countries are the most scalable existing infrastructure for closing the capacity gap. International organisations should coordinate, not duplicate. Parallel UN, OECD, Council of Europe, EU and G7 processes are already producing convergent but non-identical outputs. Format recommendations: Reduce plenary time. The current draft programme allocates over a day to plenary statements, which crowds out substantive exchange. A 60/40 split favouring multistakeholder working sessions would better serve the mandate. Adopt a Chatham House Rule track for at least one session per cluster, allowing technical experts and regulators to discuss real incidents without on-record diplomatic constraints. Mandate written outcomes per cluster, drafted by rapporteurs from across stakeholder groups, alongside the Co-Chairs' summary. Pre-circulate the Independent Scientific Panel's report at least four weeks before the session, with a structured comment period, so delegations engage with evidence rather than receive it.

Four perspectives are structurally under-represented in current AI governance discussions, beyond the general Global South framing. Smaller linguistic communities. Frontier model performance correlates with training data volume, which correlates with speaker population and digital footprint. Communities speaking languages such as Slovenian, Estonian, Maltese, or the more than two thousand African and Asian languages with minimal digital presence are governed by AI systems that do not function adequately in their own language. Inclusion requires investment in language resources and benchmarks, not only consultation. Public-sector practitioners outside capital cities. AI governance is dominated by ministries, regulators, and large NGOs. The civil servants who actually deploy AI in schools, hospitals, courts, and municipal services rarely participate. Their evidence on what works and what fails is the most policy-relevant input the Dialogue could receive. Structured practitioner panels, with travel support, would address this. Independent technical researchers outside frontier-AI countries. Most published evaluation and safety research comes from a handful of laboratories with privileged model access. Researchers in other jurisdictions cannot replicate or contest these findings. The Dialogue should advocate for structured access programmes for non-frontier academic researchers, conditional on independence from vendor interests. Affected workers and unions. Labour impact is typically discussed in macroeconomic terms. Workers directly subject to algorithmic management, automated hiring, or task displacement are rarely heard. Trade union federations and platform-worker organisations should be accredited stakeholders, not consulted afterwards.

Live evidence sessions. The Independent Scientific Panel presents a specific finding, followed by structured response from regulators, industry, and affected communities in the room. Purpose: anchor discussion to evidence rather than position. Incident clinics under Chatham House Rule. Regulators and operators discuss real AI incidents, including failures, near misses, and ambiguous cases. Purpose: surface lessons that on-record diplomatic settings suppress. Cross-stakeholder rapporteur teams. Each thematic cluster is summarised not by the Secretariat alone but by a small team drawn from a Member State, an academic, a civil-society representative, and a technical expert. Purpose: produce outputs that are credible across constituencies. Comparative obligation walkthroughs. A specific governance question, for example incident reporting, is examined across the EU AI Act, the Council of Europe Framework Convention, ISO/IEC 42001, and a national framework, side by side. Purpose: make interoperability gaps concrete rather than abstract. Capacity-pairing sessions. Capacity-holding and capacity-receiving jurisdictions meet in structured bilateral or small-group settings to identify specific, actionable cooperation. Purpose: convert the Capacity Compact from declaration into matched commitments. A common discipline applies to all five: pre-circulated material, defined questions, written outcomes. Innovative formats fail when they substitute novelty for preparation.

Regulatory frameworks: The EU AI Act provides the first comprehensive risk-based regulatory regime, with the AI Office as its enforcement spine. The Council of Europe Framework Convention on AI extends binding obligations beyond the EU, including to non-European signatories. Singapore's Model AI Governance Framework offers a lighter-touch but operational alternative suited to smaller jurisdictions. Technical evaluation infrastructure: Singapore's AI Verify Foundation provides an open-source testing toolkit, demonstrating that evaluation tooling can be a shared international good. The International Network for Advanced AI Measurement, Evaluation and Science (formerly the Network of AI Safety Institutes), bringing together members from the United States, United Kingdom, EU AI Office, Japan, Singapore, France, Canada, Republic of Korea, Kenya, India, and Australia, is developing joint evaluation protocols. MLCommons AILuminate v1.0, released in 2025, provides the first industry-standard safety benchmark across twelve hazard categories. Incident and risk monitoring: The OECD AI Incidents and Hazards Monitor (AIM) offers a working public registry, complemented by the OECD's 2025 common reporting framework for AI incidents. NIST AI RMF and ISO/IEC 42001 provide compatible vocabulary for risk management. Capacity-building: UNESCO's Readiness Assessment Methodology has been applied in over sixty countries. The EU AI Factories network now comprises nineteen factories and thirteen antennas across the EU and partner countries, including Slovenia's SLAIF, demonstrating how compute access and SME support can be operationalised at regional scale. Open ecosystems: Hugging Face and BLOOM illustrate how open development can produce credible alternatives to closed frontier systems.