Dakota State University

Success at Geneva would mean three things. First, the Dialogue produces a shared vocabulary. Right now, terms like "trustworthy AI," "accountability," and "explainability" appear in nearly every national and regional governance framework, but they mean different things in different contexts. A concrete outcome would be working definitions specific enough to serve as a baseline for future interoperability discussions — not a final answer, but a starting point that subsequent sessions can build on. Second, the Dialogue creates space for technical practitioners to speak directly to governance questions, not just governments and large organizations. The people building and deploying AI systems in clinical, educational, and public-sector contexts have empirical knowledge about where governance frameworks fail in practice. Geneva should produce a mechanism for that knowledge to flow into the process on an ongoing basis. Third, the Dialogue identifies at least one concrete gap in existing frameworks and commits to addressing it in the 2027 New York session. Broad consensus statements are useful, but they are not outcomes. If the first session ends with a specific problem named, scoped, and assigned, that would be a genuine step forward.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches
• Social, economic, ethical, cultural, linguistic and technical implications of AI

My research on EXAIM, an explainable AI middleware system for real-time multi-agent clinical decision support, surfaces a problem that cuts across all four of these areas. Most AI governance frameworks treat safety and accountability as properties of a model - something established during training and validation. But in clinical deployment, the layer between the model and the clinician carries its own governance requirements. How outputs are formatted, filtered, and explained in real time determines whether a clinician can actually exercise meaningful oversight. EXAIM addresses this by enforcing schema compliance, bounding latency, and attaching plain-language explanations to every recommendation. These are accountability mechanisms, but they operate at the interface layer, not the model layer - and current frameworks have almost nothing to say about this layer. This is why interoperability matters here too. If countries develop compatible standards for how clinical AI models are trained but no shared expectations for how outputs reach clinicians, the governance gap persists regardless of model-level alignment. The middleware layer needs to be part of interoperability discussions. The social and ethical cluster matters because the harms from poorly governed AI interfaces fall unevenly. Clinicians in under-resourced settings are less likely to have the time or training to catch AI errors that a well-designed interface would have filtered. Interface-layer governance is partly a health equity question.

The middleware layer - the software infrastructure that sits between AI models and end users - is absent from current governance frameworks and from the listed themes. This gap deserves explicit attention. Every AI system that reaches a human does so through some interface: a dashboard, an API response, a recommendation panel. That interface makes choices. It decides what to show, when to show it, how to explain it, and what to suppress. These choices have governance implications, but they are not model-level decisions. They are not captured by training data standards, model audits, or deployment licenses. They are also not easily visible to regulators. In clinical AI, this matters acutely. A model can meet every safety benchmark and still produce outputs that a poorly designed middleware layer delivers in a format that overwhelms, misleads, or goes unexplained to a clinician. The harm happens at the interface, not in the weights. The Dialogue should consider whether interface-layer governance - covering explanation requirements, output formatting standards, latency constraints, and human override mechanisms - belongs in future thematic discussions. This is not a niche technical concern. It is the layer where AI behavior and human judgment actually meet, and it is currently ungoverned across virtually every domain.

The governance gap at the interface layer has direct consequences for clinical AI deployment in the United States, particularly in under-resourced healthcare settings. The US has invested heavily in clinical AI development. Models exist for diagnostics, triage support, medication review, and clinical documentation. But deployment has been uneven, and a recurring pattern in implementation failures is not model inaccuracy — it is interface design. Clinicians receive AI outputs they cannot interpret, at volumes they cannot process, without explanations they can act on. The result is alert fatigue, ignored recommendations, and in some cases, automation bias where clinicians defer to an AI output they cannot evaluate. These are not hypothetical risks. They are documented in clinical informatics literature and in post-deployment reviews at hospital systems across the country. The governance frameworks that currently apply — FDA guidance on AI-enabled medical devices, institutional IRB review, state licensing requirements — focus almost entirely on the model. None of them specify what the interface between a clinical AI system and a clinician must be able to do. The opportunity is that this gap is addressable. The technical requirements for interface-layer governance are not prohibitively complex. Explanation generation, schema validation, latency constraints, and human override mechanisms are buildable. What is missing is the governance signal that makes building them a requirement rather than an option. For the US clinical sector, a Geneva discussion that names interface-layer governance as a priority would create pressure — on developers, on hospital procurement, and on regulators — to close a gap that currently leaves the last mile of clinical AI accountability to chance.

The AI Dialogue occupies a position no existing mechanism does: it is the only forum where every government and a genuinely diverse set of stakeholders can deliberate on AI governance without the agenda being set by the most technologically advanced economies alone. That positioning is the Dialogue's most important asset, and it should be used deliberately. The most concrete role the Dialogue can play is convergence on shared minimum standards. Not harmonization — that is probably unrealistic at this stage — but agreement on what any responsible AI governance framework must address. Interface-layer accountability, explanation requirements, and human override mechanisms are examples of things that could be specified at a minimum standard level without requiring countries to adopt identical regulatory approaches. The Dialogue can also function as an early warning system. Governance gaps that seem technical or sector-specific often generalize across contexts. The interface-layer problem I have observed in clinical AI in the US exists in different forms in educational AI systems in the Global South, in public benefits administration in Europe, and in agricultural advisory tools deployed across sub-Saharan Africa. A forum that surfaces these patterns across regions — before they become entrenched harms — is more valuable than one that responds after the fact. Finally, the Dialogue should produce something the scientific panel can work with. The Independent International Scientific Panel on AI needs governance questions to be scoped clearly enough that technical evidence can actually inform them. The Dialogue is well positioned to define those questions.

Several existing mechanisms have laid groundwork the Dialogue should connect with rather than duplicate. The OECD AI Principles and the accompanying OECD.AI Policy Observatory represent the most developed cross-national framework for AI governance to date. The Observatory's policy tracking across 70+ countries is a resource the Dialogue should draw on directly when identifying convergence points and gaps. The limitation is that OECD membership skews toward wealthier nations — the Dialogue's universal mandate can extend that work to contexts the Observatory does not adequately cover. UNESCO's Recommendation on the Ethics of Artificial Intelligence, adopted by 194 member states, establishes shared ethical grounding. The Dialogue should treat it as a floor, not a ceiling, and focus on translating its principles into operational requirements that national regulators can actually implement. The ITU's AI for Good initiative, which hosts the Geneva summit alongside the Dialogue's first session, is a natural technical partner. Its work on AI standards and capacity building in developing countries addresses the access dimension that pure governance frameworks often miss. What the Dialogue adds that none of these provide is a space where governments and non-state stakeholders deliberate as peers on normative questions — not just share best practices. The OECD is a club. UNESCO produces recommendations. ITU focuses on technical standards. The Dialogue is the only mechanism with the mandate to ask what global AI governance should require, from everyone, and to work toward answers that reflect more than the preferences of the most powerful actors in the room.

The current structure risks reproducing the same dynamic it is trying to avoid: governments and large organizations dominating the room while technical practitioners, civil society, and researchers from smaller institutions contribute written inputs that may or may not influence what happens in plenary. A few structural changes would help. Thematic working sessions should be organized around specific governance problems, not broad topic areas. "Safe and trustworthy AI" as a session title invites general statements. "What should an AI system be required to explain to a clinician before a treatment decision" invites people with actual knowledge of the problem to contribute something useful. Problem-scoped sessions naturally draw the right stakeholders into the room. Written inputs like this one should feed visibly into session design. If a recurring theme across submissions is the interface-layer governance gap, that theme should appear in the Geneva agenda with attribution to the submissions that raised it. Stakeholders need to see that written contributions shape outcomes, not just inform them. Finally, the Dialogue should create a dedicated track for early-career researchers and practitioners. Not a side event — a track with real agenda time. The people building and deploying AI systems right now are often early in their careers. Their empirical knowledge of where governance frameworks fail in practice is not well represented in intergovernmental forums, and the Dialogue has an opportunity to change that.

Several communities are consistently underrepresented in global AI governance discussions, and the patterns are not accidental. Clinicians, teachers, social workers, and other frontline professionals who interact with AI systems daily are almost entirely absent. These are the people who experience governance gaps directly — who receive AI outputs they cannot interpret, who absorb the consequences of poorly designed interfaces, who develop workarounds that never get documented. Their knowledge is practical and specific in ways that policy discussions rarely are. Including them requires going beyond written input calls and creating structured ways for practitioner testimony to enter the record. Researchers from institutions outside major research universities face a different barrier. The cost of participation — travel, registration, time — effectively excludes researchers from smaller institutions, community colleges, and universities in lower-income countries who are doing relevant applied work. Fellowship or sponsorship mechanisms tied to the Dialogue's Geneva and New York sessions would help, but only if they are designed to reach people who would not otherwise apply. Communities directly affected by AI deployment in high-stakes contexts — patients whose care is influenced by clinical AI, students subject to automated assessment, people whose benefits eligibility is determined algorithmically — have no formal pathway into these discussions. Their inclusion is not just a fairness question. Governance frameworks that are not informed by the experience of people on the receiving end of AI decisions will continue to miss the harms that matter most.

Three formats worth considering. Structured red-teaming sessions, where small mixed groups — a regulator, a technical practitioner, a frontline professional, a civil society representative — are given a specific governance proposal and asked to identify how it would fail in their context. This is more generative than panel discussions and produces concrete feedback that session chairs can act on. It also forces stakeholders who rarely talk to each other to engage with the same problem from different angles. Longitudinal case submissions, where stakeholders submit not just position statements but documented examples of AI governance working or failing in practice, with follow-up sessions that interrogate those cases in depth. A single well-documented case of interface-layer failure in a clinical setting tells the Dialogue more about what governance needs to address than a dozen general recommendations. Building a shared case library across sessions would give the Dialogue an evidence base that accumulates over time. Asynchronous deliberation tracks that run between sessions. The gap between Geneva 2026 and New York 2027 is nearly a year. A facilitated online process — not just a mailing list, but structured deliberation with clear inputs and outputs — would allow the work to continue between in-person meetings and ensure that the New York session builds on Geneva rather than partially repeating it. This would also allow participation from stakeholders who cannot attend in person, which is most of the world.

The most instructive examples of effective AI governance are not the headline frameworks - they are the specific implementation decisions that made abstract principles operational. The EU AI Act's risk-tiering approach is worth building on, not because it is complete, but because it established a replicable logic: the governance requirements on an AI system should scale with the stakes of the decisions it influences. Applying that logic to the interface layer would mean that a clinical AI system presenting recommendations to a physician in real time faces stricter explanation and override requirements than a low-stakes recommendation engine. The tiering concept is sound. It just has not been extended far enough down the deployment stack. At the system level, EXAIM - an explainable AI middleware architecture I developed and published at IEEE ICHI 2026 - offers a concrete example of what interface-layer governance looks like in practice. The system enforces schema compliance across AI agent outputs, bounds the rate at which updates reach clinicians, and generates plain-language explanations for every recommendation before it reaches a human. It achieved 96.8% schema compliance, 1.22 second median latency, and a 75% reduction in clinician-facing update frequency in testing. These outcomes were not accidental - they resulted from treating the interface between AI and human judgment as a governed layer with explicit requirements, not an implementation detail. The broader lesson is that governance works when it specifies what a system must be able to do at the point of human contact, not just how it was built. Frameworks that stop at the model leave the most consequential decisions - what gets shown, when, and how - entirely to developers.