Independent Researcher / Taiwan

Success will mean the first global dialogue producing a shared language and governance direction that goes beyond being limited to model-side safety. Three outcomes will make this dialogue truly significant. First, the dialogue should establish a set of agreed-upon terms to distinguish different types of AI-related harms, encompassing not only errors generated by the system but also user-side risks arising from prolonged interactive dialogues. Without such distinctions, governance frameworks will continue to treat structurally different issues as the same category. Second, the dialogue should explicitly recognize public AI literacy and calibration as governance infrastructure, rather than merely as supplementary education. If only the channels of AI usage are addressed without involving the interpretation and decision-making conditions of ordinary users, the capacity-building agenda will be incomplete. Third, the dialogue should reach a consensus acknowledging that independent public-interest research on AI interaction risks is currently limited by asymmetric access to interactive data. Viewing this as a structural barrier and advocating for the inclusion of de-identified research access under an ethical governance framework, as part of the interoperability agenda, will constitute meaningful and feasible progress. A dialogue achieving these three outcomes will shift AI governance from frameworks primarily focused on vendor obligations and system benchmarks toward governance systems that also consider the social conditions for the safe use of AI.

• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI

One issue not captured by the listed themes is the structural gap in research access to AI interaction data. Most governance discussions assume that evidence about AI-related harms can be independently examined and verified. In practice, the conversation data needed to study user-side risks, including patterns of miscalibration, escalating reliance, and contextual drift across long interactions, sits almost entirely within commercial platforms. Independent researchers and public interest organizations have no structured, ethically governed pathway to access de-identified records for this purpose. This means that governance language for user-side interaction risks can move faster than the evidence that should inform and constrain it. That is not a good position for governance to be in. The problem does not belong cleanly to any single theme on the list. It touches capacity-building, interoperability, and transparency at the same time. It also affects the quality of evidence available to policymakers when they assess claims about social harms from AI use. The Global Dialogue should consider naming this barrier directly and recognizing that facilitating ethically governed, de-identified research access to interaction data, with appropriate privacy protections and institutional safeguards, is part of its broader agenda rather than a peripheral technical matter. Closing this gap would strengthen the evidence foundation for every other theme the Dialogue is trying to address.

One issue not captured by the listed themes is the structural gap in research access to AI interaction data. Most governance discussions assume that evidence about AI-related harms can be independently examined and verified. In practice, the conversation data needed to study user-side risks, including patterns of miscalibration, escalating reliance, and contextual drift across long interactions, sits almost entirely within commercial platforms. Independent researchers and public interest organizations have no structured, ethically governed pathway to access de-identified records for this purpose. This means that governance language for user-side interaction risks can move faster than the evidence that should inform and constrain it. That is not a good position for governance to be in. The problem does not belong cleanly to any single theme on the list. It touches capacity-building, interoperability, and transparency at the same time. It also affects the quality of evidence available to policymakers when they assess claims about social harms from AI use. The Global Dialogue should consider naming this barrier directly and recognizing that facilitating ethically governed, de-identified research access to interaction data, with appropriate privacy protections and institutional safeguards, is part of its broader agenda rather than a peripheral technical matter. Closing this gap would strengthen the evidence foundation for every other theme the Dialogue is trying to address.

Taiwan sits in an instructive position for this discussion. It has relatively high internet penetration and active AI adoption across both public and private sectors, but like most places it has not yet developed public-facing governance infrastructure for the specific risks that arise from extended conversational AI use. The most significant challenge is not access. It is interpretive capacity. Users across education, healthcare adjacent contexts, and everyday decision-making are increasingly relying on conversational AI in ways that involve trust, emotional engagement, and repeated interaction over time. The governance frameworks currently available do not address what happens in those settings when users miscalibrate capability, misunderstand memory behavior, or gradually normalize unstable interaction patterns. There is also a regional dimension. Much of the governance standard-setting for AI is happening in jurisdictions with larger regulatory infrastructure. Smaller economies and civil society actors in the Asia-Pacific region often engage these standards as recipients rather than contributors. A Global Dialogue that creates genuine space for independent research contributions, regardless of institutional affiliation or geographic origin, helps address that imbalance. The opportunity is that Taiwan and similar contexts can offer concrete, grounded observations about how ordinary users actually engage conversational AI in daily life, which is still an underrepresented perspective in international governance discussions. Research coming from outside the major platform ecosystems and outside large institutional frameworks can sometimes see things that internal teams and large-country regulators are slower to name. What is missing is a shared governance vocabulary for these observations to connect to, and structured pathways for independent researchers to contribute evidence rather than only commentary. Both gaps are addressable within the themes this Dialogue has already identified.

The most useful role the Dialogue can play is producing shared language that travels across jurisdictions and institutional contexts without requiring uniform regulation. International cooperation on AI governance has stalled in places not because of disagreement on values but because the vocabulary is inconsistent. Different frameworks use the same words to mean different things, or use different words for the same problem. A Dialogue that produces clearer shared definitions, even provisional ones, gives subsequent regional and national processes something to build from rather than restart. Specifically, the Dialogue can help establish that AI governance encompasses not only how systems are designed and deployed but also the conditions under which people understand and rely on them. That shift in framing has not yet been made explicit in most international documents. Making it explicit would give capacity-building, transparency, and safety discussions a more complete foundation. The Dialogue can also play a convening role that formal treaty processes cannot. It can surface perspectives from independent researchers, civil society actors, and smaller economies that are often present in consultation processes but rarely shape the output. If the Dialogue takes those contributions seriously enough to reflect them in its Co-Chairs summary, it will have demonstrated a model of inclusive governance that other processes can follow.

Several existing frameworks have laid important groundwork. UNESCO's Recommendation on the Ethics of Artificial Intelligence established a broad normative foundation covering education, public understanding, and individual autonomy. The OECD AI Principles introduced risk-based thinking and have been adopted across a wide range of national contexts. The EU AI Act has moved literacy and transparency from aspiration into binding regulatory language. The Council of Europe Framework Convention has grounded AI governance in human rights and the rule of law. The Global Digital Compact created a cross-cutting commitment to safe and trustworthy AI at the UN level. These are strong foundations. The gap is not in values but in operational specificity for certain categories of risk, particularly those that emerge through ordinary public use of conversational AI rather than through high-stakes institutional deployment. The added value the Dialogue can bring is synthesis and extension. It can connect the literacy strand in UNESCO and OECD work with the transparency and accountability strand in the EU and Council of Europe instruments, and do so in language accessible to Member States without advanced regulatory infrastructure. It can also create a space where independent research contributions, not only contributions from governments and large organizations, inform the shared record. One concrete opportunity is to build on the OECD and UNESCO literacy work by extending it explicitly to conversational AI settings, where the risks of miscalibrated reliance and long-context drift are not yet covered by existing guidance. That extension does not require new institutions. It requires a deliberate signal from this Dialogue that the issue belongs on the agenda.

The Dialogue should distinguish between stakeholders who set direction and stakeholders who contribute evidence. Both matter, but they need different participation structures. Governments and intergovernmental bodies are the appropriate decision-makers for binding commitments and formal outcomes. Their participation should remain central. But the quality of those decisions depends heavily on evidence and perspective that governments alone do not hold. Independent researchers, civil society practitioners, and users with direct experience of AI-related harms carry knowledge that is not well captured by formal consultation rounds or written inputs alone. A more effective structure would create tiered participation: a plenary process for governments and major institutional stakeholders, and a parallel track with lighter entry requirements for independent researchers, smaller civil society organizations, and practitioners from underrepresented regions. Written inputs are a good start, but they tend to produce documents that are read selectively. Shorter structured submissions with clear thematic tagging, combined with small facilitated sessions where contributors can respond to questions, would produce more usable evidence for the Co-Chairs. The Dialogue should also commit to publishing all written contributions in full, with searchable access, rather than only summarizing them. A searchable public record creates a commons that later governance processes can draw from.

The most underrepresented perspectives are ordinary users, independent researchers without institutional affiliation, and practitioners from smaller economies in the Global South and Asia-Pacific region. Ordinary users are the people most directly affected by the risks this Dialogue is meant to address, but governance processes have no reliable mechanism for capturing their experience. User experience research, complaint data, and civil society case documentation are the closest proxies, but they are rarely treated as primary evidence in intergovernmental settings. Independent researchers face a structural barrier. Without university affiliation, foundation backing, or NGO credentials, it is difficult to participate in formal consultation processes even when the research is substantive and publicly available. The written inputs process partially addresses this, but there is no pathway from written input to actual engagement with the process. Smaller economies face a different problem. They often have governance priorities shaped by specific social and cultural conditions of AI adoption that do not map onto the high-stakes deployment scenarios that dominate international discussions. A companionship AI risk looks different in a context with limited mental health infrastructure than it does in a jurisdiction with robust clinical oversight. Those differences matter for governance and are currently invisible in most international documents. Including these voices requires deliberate structural choices: accessible submission formats, interpretation support, geographic diversity requirements for any expert panels, and a genuine commitment to reflecting non-institutional perspectives in official outputs.

The most effective format change would be replacing large plenary panels with smaller, problem-focused working sessions organized around specific governance gaps rather than broad themes. Large panels produce statements. Small working sessions produce distinctions. The Dialogue needs more of the latter. A session organized around a single question, such as what transparency requirements for conversational AI memory behavior should look like, will generate more usable output than a panel on trustworthy AI in general. Asynchronous participation options would also significantly expand the contributor base. Not everyone with relevant expertise can travel to New York or Geneva or participate in real-time sessions across time zones. A structured asynchronous track, with written exchanges that feed into facilitated summaries, would allow practitioners and researchers from underrepresented regions to contribute substantively without the participation barriers that currently exclude them. Finally, the Dialogue should consider publishing interim synthesis documents during the process, not only at the end. If contributors can see how their inputs are being interpreted and respond before the final summary is produced, the output will more accurately reflect the actual range of perspectives submitted. That feedback loop is standard in good deliberative processes and mostly absent from intergovernmental ones.

The most underrepresented perspectives are ordinary users, independent researchers without institutional affiliation, and practitioners from smaller economies in the Global South and Asia-Pacific region. Ordinary users are the people most directly affected by the risks this Dialogue is meant to address, but governance processes have no reliable mechanism for capturing their experience. User experience research, complaint data, and civil society case documentation are the closest proxies, but they are rarely treated as primary evidence in intergovernmental settings. Independent researchers face a structural barrier. Without university affiliation, foundation backing, or NGO credentials, it is difficult to participate in formal consultation processes even when the research is substantive and publicly available. The written inputs process partially addresses this, but there is no pathway from written input to actual engagement with the process. There is a deeper issue worth naming. The scale of change that AI is introducing means that established institutions and disciplinary frameworks are often poorly positioned to see what is actually happening. What is needed is not only more expertise in the conventional sense. It is more non-linear, systems-level thinking from people whose backgrounds sit outside traditional credentialing pathways. Voices shaped by different cultural contexts, unconventional educational trajectories, and direct lived experience with these technologies often recognize structural problems that specialists trained within existing paradigms are slower to identify. Excluding those voices is not a neutral omission. It is one reason governance development has lagged behind technical development. Including underrepresented perspectives requires deliberate structural choices: accessible submission formats, interpretation support, geographic diversity requirements for expert panels, and a genuine commitment to reflecting non-institutional perspectives in official outputs.