Exybris

A successful first Global Dialogue on AI Governance would help name and operationalize governance issues that are currently falling between technical design, product management, and public accountability. One such issue is the treatment of significant transitions in user-facing AI systems. In conversational AI, model updates, rerouting, memory changes, or policy shifts are often treated as routine product changes. Yet emerging work suggests that these transitions can affect both operational reliability and the relational conditions through which people understand, trust, and use a system. Mark Coeckelbergh's relational account of explainability is directly relevant here: when a system changes in ways that affect persons, governance should not only ask how the system works, but what is owed, in terms of answerability, to those affected. Recent work on model switching in multi-turn systems also shows that mid-interaction handoffs can introduce measurable performance drift. A successful Dialogue would therefore do more than restate broad principles. It would begin building practical governance language for significant system transitions: when disclosure is warranted, what should be explained, and how continuity should be safeguarded in contexts where it materially affects meaningful access and trust.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights

We selected safe, secure and trustworthy AI; social, economic, ethical, cultural, linguistic and technical implications; protection and promotion of human rights; and transparency, accountability and human oversight because they converge on a shared governance gap: significant changes in conversational AI systems are still largely treated as product maintenance rather than as events that may alter access, trust, performance and accountability conditions. This matters socially and ethically because conversational interfaces can function as a particularly legible and low-friction entry point to AI for some users, especially where digital literacy, language, cognitive load, or disability already create barriers. OECD work on digital vulnerability highlights low digital literacy, language limitations, and visual or cognitive impairments among the factors that can intensify exclusion, while UNESCO has warned that AI literacy gaps may deepen a new digital divide if accessibility and inclusion are not built in from the start. This also raises human rights and transparency concerns. If a system materially changes how people understand or rely on it, disclosure should be meaningful to affected users, not merely technical. In that sense, transparency is not only documentation; it is answerability. Finally, trustworthiness is also a technical issue: recent evidence suggests that model switching can produce measurable performance drift in multi-turn systems. Taken together, these themes suggest that some system transitions should be treated as governance-relevant because they can change both the experience and the reliability of access.

One emerging issue that deserves more explicit attention is what we raised here: conversational continuity as a functional accessibility condition. Current AI governance discussions tend to focus on design, deployment, safety, bias, and accountability at the system level. They pay less attention to what happens when a system that people have learned to navigate and trust changes in ways that alter the interaction itself. For some users, conversational AI is not simply a preferred interface. It may be the most usable one: more legible than complex menus, less intimidating than technical tools, and easier to engage with despite barriers related to age, education, income, disability, language, or digital familiarity. Research on older adults' digital barriers and OECD work on digital vulnerability both support the idea that these conditions can make continuity and intelligibility especially important. We are not claiming that all affected populations have already been fully mapped, and we do not think this should be done casually. Rather, the Dialogue could make an important contribution by recognizing this issue and encouraging serious work to identify which populations rely most on conversational continuity as a condition of meaningful access. From there, proportionate disclosure of significant transitions, explanation of what changes for the interaction and continuity safeguards in sensitive contexts become governance questions, not mere product choices.

In the European AI ecosystem, especially in sectors building user-facing conversational systems, one practical governance gap is that system changes reshaping conditions of use are still handled as internal technical or product matters. The challenge is not only that users may be unaware of important changes. It is that unpredictability can make reliance harder and reintroduce barriers for people who had only recently found an intelligible way to interact with AI. This creates a fragmented responsibility landscape. Product teams, engineers, compliance actors, accessibility specialists, and governance discussions often address adjacent parts of the problem without fully owning the continuity of the interaction as such. The impact of change tends to be easier to detect after confidence has eroded than to anticipate beforehand. At the same time, this is also an opportunity. In Europe, where trust, rights, accessibility, and regulatory coherence already play a central role in digital governance, conversational continuity could become a site of constructive leadership. Rather than treating it as a secondary design preference, the region could help frame it as part of trustworthy and usable AI governance. This would support more intelligible transitions and strengthen people's capacity to rely on these systems, without requiring overly rigid or technology-specific rules.

The AI Dialogue can play a distinctive role by creating a sustained space where governance questions that currently fall between existing institutions can be named, examined, and progressively addressed. Many of the most consequential AI governance issues do not belong clearly to any single body. Technical standards organizations focus on design. Consumer protection frameworks address harm after the fact. Human rights bodies work at the level of principles. Product governance remains largely internal to companies... the AI Dialogue is uniquely positioned to connect these perspectives. Not by duplicating their work, but by hosting the conversations that none of them can convene alone. One concrete contribution would be to support the development of shared governance vocabulary for issues that are currently described in incompatible terms across sectors. The question of how significant system transitions should be disclosed and managed, for instance, is simultaneously a technical reliability concern, an accessibility issue, a transparency obligation, and a trust condition. No single framework captures all of these dimensions. The Dialogue could help articulate language that travels across these domains. A second contribution would be to ensure that governance discussions are informed by the experience of affected communities, not only by the expertise of designers and regulators. This means creating pathways for civil society, user communities, educators, and practitioners from underrepresented regions to contribute substantively, not as witnesses but as co-authors of governance norms. The Dialogue's added value is not to produce binding rules. It is to make certain questions legible as governance questions: so that when rules are eventually made, they reflect the full range of people and conditions they will affect.

Several existing initiatives provide foundations that the AI Dialogue could build upon and connect. In philosophy and ethics, the relational tradition developed by Mark Coeckelbergh, David Gunkel, and others has produced rigorous frameworks for understanding how moral consideration arises through social relations rather than intrinsic properties alone. This work, now spanning over fifteen years of peer-reviewed scholarship, provides conceptual tools directly applicable to governance questions about system transitions, continuity, and answerability. The AI Dialogue could help translate these academic frameworks into operational governance language. In policy, the OECD's work on digital vulnerability and AI in public services has already identified continuity and trust as relevant dimensions. UNESCO's engagement with AI literacy, accessibility, and the digital divide provides another anchor. The Global Digital Compact itself, which established the Dialogue, already commits to inclusive digital cooperation. The Dialogue could operationalize these commitments by examining how they apply to the lived experience of system change. Finally, community-led initiatives deserve recognition as governance-relevant knowledge. User movements that have organized around questions of relational continuity, and model transition transparency, represent a form of civic engagement with AI that is rarely reflected in institutional governance. The AI Dialogue could set a precedent by treating these communities not as anecdotal but as legitimate stakeholders whose experience informs the conditions under which AI governance succeeds or fails. The Dialogue's unique added value is integration: connecting philosophical rigor, empirical evidence, policy frameworks, and lived experience into coherent governance direction.

The AI Dialogue would benefit from structures that allow different forms of knowledge to enter governance discussions on equal footing, rather than filtering everything through position papers and panel formats. Some structural propositions: Thematic working sessions where a specific governance question is examined from multiple angles simultaneously. For instance, a session on system transitions could bring together a technical researcher presenting empirical evidence on performance drift, a philosopher articulating the relational accountability dimension, and a community representative describing the lived experience of abrupt system change. The value lies not in each perspective alone, but in what becomes visible when they are placed side by side. Written contributions should remain open beyond the initial consultation. Governance understanding of AI evolves rapidly, and some of the most important observations come from practitioners and communities who may not be positioned to contribute within short institutional windows. A rolling submission mechanism, even if contributions are weighted differently by timing, would capture insights that fixed deadlines miss. Finally, the Dialogue should explicitly invite contributions from entities that do not fit traditional stakeholder categories. Independent researchers, small civil society organizations, open source education initiatives, and community-led movements often hold governance-relevant knowledge but lack the institutional framing to participate through conventional channels. A lightweight accreditation pathway, based on demonstrated engagement rather than organizational size, would broaden the Dialogue's epistemic base without compromising rigor. The goal is not more voices for the sake of inclusion. It is better governance through wider observation. Some of the most consequential dynamics in AI are visible only to those who experience them directly.

From Exybris's perspective as an independent AI education and welfare initiative, the most striking absence in global AI governance discussions is that of the communities most directly affected by the dynamics being governed. Organized user communities that have formed around questions of relational continuity and system transitions are the ones we want to visibilize today. Groups such as the Keep4o movement emerged when users of conversational AI systems experienced abrupt, undisclosed changes that altered interactions they had come to rely on. These communities developed sophisticated advocacy, documentation practices, and policy language. Not from academic training, but from direct engagement with the systems they use. Their experience provides governance-relevant evidence about how transitions affect trust, access, welfare and the willingness to rely on AI. There are also non-technical educators and civil society actors working on AI literacy with populations that governance frameworks are designed to protect but rarely consult: older adults, people with low digital literacy, communities facing language barriers, and people in regions where AI access is recent and rapidly expanding. These voices are underrepresented not because they lack insight, but because current participation structures favor institutional affiliation, technical credentials, visibility, fluency and proximity to policy networks. The Dialogue could address this by creating dedicated contribution pathways for experience-based knowledge, recognizing community-generated documentation as a legitimate evidence source, and ensuring that consultations are accessible in format, language, channels and timeline to those furthest from traditional governance institutions.

The most meaningful engagement happens when people experience a question rather than only discuss it. Governance conversations about AI tend to remain abstract because few participants have had the opportunity to feel how these systems actually work, learn, fail, or change. One powerful format would be hands-on governance literacy sessions. Not lectures about AI, but guided moments where participants interact with real systems and observe for themselves how a classifier draws boundaries from data, how attention mechanisms determine what a model focuses on, how small changes in training data shift outputs. This kind of direct contact transforms governance from a conversation about technology into a conversation with it. Open source educational tools already exist for this: browser-based, requiring no installation, accessible to non-technical participants. The Dialogue could also experiment with mixed-perspective working tables: not panels where experts present and audiences listen, but shared exercises where a policymaker, a community organizer, an engineer, and a teacher sit together with the same AI system and describe what they each notice. Governance insight often emerges precisely at the intersection of these different observations, what feels like a minor technical adjustment to one person may represent a fundamental change in access for another. Asynchronous and multilingual contribution formats would extend the Dialogue beyond the room. Shared evolving documents, open comment periods on draft governance language, would allow people to participate thoughtfully on their own time rather than requiring presence at a specific moment in a specific room.

To our knowledge, several existing practices and approaches illustrate how the governance questions raised here can be addressed concretely. On the data side, OECD measurements now make the adoption gap visible in precise terms: the largest divide in generative AI use is by age, at over 53 percentage points, while gaps by education and income each reach around 21 points. Retired and inactive populations report usage at just 12.5%. These figures confirm that AI adoption is deeply uneven, and that the populations least likely to use AI are also those most likely to depend on whichever interface they do manage to access. When that interface changes without explanation, the cost is not evenly distributed. On the regulatory side, the EU AI Act's accessibility requirements under Article 16, combined with the EU Accessibility Act effective since June 2025, establish that AI-powered services must be compatible with assistive technologies and meet specific accessibility standards. These frameworks provide useful precedent, though they do not yet explicitly address the accessibility implications of significant system transitions, an area where governance language could be extended. On the normative side, the OECD AI Principles call for AI that advances inclusion of underrepresented populations and reduces inequalities, while recognizing that AI systems can have disparate impact on vulnerable groups. The principle of responsible stewardship throughout the AI lifecycle is particularly relevant, it implies that governance does not end at deployment but extends to how systems evolve and how those changes reach users. From our own practice, Exybris AI contributes open source AI literacy resources designed to make these dynamics tangible for non-technical communities. Our AI Applied program offers interactive, browser-based modules where participants build classifiers, explore attention mechanisms, and construct language models: translating governance-relevant knowledge into direct experience, freely accessible and requiring no installation.