Holistic Research Canada

Success would mean more than a communiqué. It would mean the Dialogue produced something structurally durable — shared governance infrastructure that doesn't dissolve when delegations return home. Three concrete outcomes would mark genuine success. First, agreement on minimum standards for health and human services AI, including transparent accountability for automated decision-making that affects access to care. These are not aspirational principles — they are floor-level requirements that millions of people already need. Second, a mechanism for ongoing participatory review: a standing process through which civil society, frontline health workers, and communities affected by AI systems can meaningfully revise governance frameworks as technology evolves. One-time declarations age badly; iterative processes don't. Third, recognition that data governance and AI governance are inseparable. No AI dialogue that ignores who owns health data, who can link it across systems, and whose clinical knowledge gets encoded in training datasets will produce governance that holds. Across all three, the Dialogue would mark a genuine advance if it moved beyond nation-state-only representation and formally included sub-national governments, Indigenous data sovereignty bodies, and sectoral regulators who carry actual implementation responsibility. In federated countries like Canada, national commitments to AI governance mean very little if provincial health ministries and Indigenous communities aren't at the table designing the mechanisms. Finally, success would mean the Dialogue produced a shared vocabulary for harm — specific, measurable, and clinically grounded — rather than the vague risk language that currently dominates international AI frameworks. Governance that cannot describe harm with precision cannot prevent it.

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

These four priorities are not abstract governance ideals - they describe the specific conditions under which health AI either serves or harms the people it claims to help. Safe, secure and trustworthy AI is foundational for health and human services contexts, where automated or AI-assisted decisions affect access to care, benefit eligibility, and clinical treatment. Safety in these domains is not primarily a cybersecurity question; it is a clinical and ethical one. Trustworthy AI requires that systems perform equitably across populations, that failure modes are identified before deployment rather than after harm, and that affected communities have genuine recourse when systems fail them. Interoperability of governance approaches is an urgent priority for federated countries and cross-border health systems. In Canada, the fragmentation of health data governance across federal, provincial, and Indigenous jurisdictions means that national AI governance commitments routinely fail at the implementation layer. Internationally, divergent governance frameworks create conditions where AI systems developed under strong regulatory environments are deployed without equivalent protections in lower-resource settings. Interoperability is the mechanism through which governance commitments become mutually reinforcing rather than mutually undermining. Protection and promotion of human rights anchors AI governance in existing legal obligations that states have already accepted. Health AI raises specific human rights concerns - informed consent, non-discrimination, privacy, and the right to explanation - that must be named explicitly rather than subsumed under general risk language. Transparency, accountability, and human oversight are the operational requirements that make the other three priorities enforceable. Governance frameworks without transparency mechanisms are unverifiable. Accountability without clear lines of responsibility is ceremonial. Human oversight, particularly by clinicians and affected communities, is the only reliable check on the gap between how AI systems are designed and how they actually function in practice.

Three issues warrant explicit attention that the current thematic framework does not adequately surface. The governance of AI training data as a distinct domain. Current AI governance frameworks focus heavily on deployment - how AI systems behave once released. Far less attention is paid to the governance of the data used to build them. In health contexts, this gap is consequential: whose clinical presentations are represented in training datasets, which languages and cultural contexts are encoded as normative, and which populations are systematically underrepresented determines the equity ceiling of any AI system before a single governance policy is applied. Training data governance requires its own framework, including provenance requirements, consent standards, and equity audits at the point of dataset construction - not only at the point of deployment. The compounding effect of AI governance gaps on already-fragmented federated systems. International AI governance discourse is predominantly designed around unitary state actors. Federated countries - where health, education, and social services are constitutionally sub-national responsibilities - face a structural mismatch between where governance authority sits and where AI governance commitments are made. This is not a national implementation detail; it is a design flaw in current governance architecture that reliably produces accountability gaps at the jurisdictions closest to affected populations. The epistemic dimension of AI governance: whose knowledge gets built in. AI systems in health and human services encode assumptions about what constitutes evidence, which outcomes matter, and which forms of knowing are valid. Indigenous, community, and patient knowledge are systematically excluded from these encoding decisions. This is not a representation problem solvable by adding diverse voices to advisory panels - it is a structural question about the epistemological foundations of AI systems that governance frameworks have not yet named, let alone addressed.

Canada's federated structure creates a compounding governance gap in health AI that is rarely named clearly in international forums. Jurisdiction over health is provincial; AI development is predominantly federal or private-sector; data sovereignty for Indigenous Peoples sits in a third distinct domain. The result is that no single actor has both the authority and the information needed to govern health AI coherently. In practice, this means clinical AI tools are being adopted in mental health, primary care, and hospital systems without agreed minimum standards for outcome monitoring, bias evaluation, or transparency to patients. The Clinical Data Governance Checklist (CDGC; Hansen, 2026) — a measurement instrument developed specifically to address this — identifies twelve governance dimensions where Canadian health systems consistently score lowest: federated data linkage protocols, equity-stratified outcome reporting, and patient-facing algorithmic transparency. These aren't edge cases; they describe the current norm. The harm is concrete. Outcome monitoring systems in mental health — including feedback-informed treatment approaches — depend on data that can be safely aggregated, compared across providers, and reported to funders without exposing vulnerable patient populations. When governance frameworks are absent or fragmented, organizations default to the least risky option: no data sharing at all. Equity suffers first. Indigenous, rural, and low-income populations are consistently excluded from normative datasets, meaning AI tools calibrated on majority populations perform worst for those with greatest need. The opportunity is equally concrete. Canada's scale and linguistic diversity make it a meaningful testbed for federated AI governance — models that preserve provincial authority while enabling cross-jurisdictional learning. The CIRANO-led work on federal-provincial-territorial data interoperability points toward what's possible. The gap is not technical. It is governance will, shared vocabulary, and accountable process.

The AI Dialogue's most valuable contribution would be bridging the growing distance between high-level AI governance frameworks — which tend to be abstract, aspirational, and produced at speed — and the sectoral, sub-national, and community-level governance realities where AI is actually deployed and experienced. International AI governance currently suffers from a translation problem. Principles exist in abundance. What is missing is a shared methodology for moving from principle to practice — particularly in health, social services, and public administration, where the populations most affected by AI decisions are also least represented in governance design processes. The Dialogue could serve as the first standing forum where this translation work happens visibly and with accountability. Specifically, it could establish sector-specific working groups that develop implementable governance guidance for high-stakes application domains: clinical AI, predictive policing, social benefit administration, and education. Each working group would include sub-national regulators, affected community representatives, and domain clinicians or practitioners — not only AI developers and national delegates. The Dialogue could also play a critical norm-setting role for smaller states and under-resourced governments that lack the technical capacity to evaluate AI governance frameworks independently. By producing practical, modular governance tools — checklists, assessment rubrics, minimum disclosure standards — it could give implementers something they can actually use rather than principles they can only affirm. Perhaps most importantly, the Dialogue could make visible the cumulative effect of national AI governance decisions on global health equity. When wealthy countries adopt AI systems that encode their population's clinical norms as universal, the downstream effects on global South health systems are significant and largely invisible in current governance discourse. A standing Dialogue creates the accountability structure to name and address this.

Several existing initiatives offer genuine architecture that the Dialogue should connect with rather than duplicate. The CIRANO work on federal-provincial-territorial data governance in Canada — particularly the analysis by Dudoit, LaBillois, and Oliveira on AI policy coherence across FPT systems — offers a worked example of federated governance design with real implementation history. It deserves more international attention than it receives. The OECD AI Policy Observatory provides the most comprehensive comparative database of national AI governance approaches, and the Dialogue should use it as a living evidence base rather than starting from scratch. The OECD's work on AI and health data — including standards for algorithmic transparency in clinical systems — is particularly relevant. UNESCO's Recommendation on the Ethics of AI (2021) remains the broadest multilateral framework with explicit attention to human rights and cultural diversity. The Dialogue should treat it as a floor, not a ceiling, and focus its energy on implementation rather than re-negotiating principles that have already been agreed. For Indigenous data governance specifically, the CARE Principles for Indigenous Data Governance and the work of the Global Indigenous Data Alliance provide frameworks the Dialogue should formally incorporate. The OCAP® principles developed by the First Nations Information Governance Centre in Canada represent a mature, field-tested model for community-controlled data sovereignty that could be adapted internationally. Finally, the Global Partnership on AI (GPAI) working groups on responsible AI and data governance have produced technically rigorous output that often goes under-referenced in policy-level discussions. The Dialogue would benefit from formal linkage with GPAI's sectoral work rather than maintaining parallel but disconnected tracks. The added value the Dialogue can bring is convening authority: the ability to bring these streams into explicit relationship and produce governance commitments that are cross-referenced, not contradictory.

Effective multi-stakeholder contribution requires structural design, not good intentions. Too many international AI dialogues include diverse stakeholders in formats that make substantive contribution impossible — large plenary panels, pre-determined outcomes, and consultation windows that open after key decisions have been made. The Dialogue should organize stakeholder contribution around function rather than identity. Civil society organizations don't need a seat at the table to be seen; they need a seat in the drafting process. Health and education sector practitioners don't need to present case studies in plenary; they need to co-author governance criteria with technical experts. Indigenous communities and data sovereignty bodies don't need a dedicated panel; they need co-design authority over the governance provisions that affect their data. Structurally, I would recommend: pre-Dialogue deliberative process where affected communities develop governance priorities independently before engaging with state actors; a tiered working group structure with clear scope and decision rights; and a post-Dialogue implementation monitoring mechanism with civil society reporting capacity. For format: smaller structured working sessions consistently outperform large panels for substantive progress. Breakout groups with diverse composition, a clear question to answer, and a reporting mechanism produce more usable output than keynote-and-reaction formats. Simultaneous interpretation and accessible participation options (hybrid attendance, asynchronous input) are minimum requirements for genuine inclusion, not optional enhancements. For civil society, academic researchers, and health sector contributors specifically: ensure submissions are actually read, cited, and traceable in outcome documents. The credibility of the Dialogue depends on whether participant contributions are visibly incorporated or ceremonially received and ignored. The Dialogue should also formally recognize the contribution of practitioners in low-resource settings who carry implementation knowledge that is irreplaceable — and compensate their participation, rather than requiring self-funding as a condition of inclusion.

The most consequential absences in global AI governance discussions are not random. They follow the same patterns as every prior technology governance process: the people most affected by automated decisions are least represented in designing the systems that govern them. In health AI specifically: patients with lived experience of mental health conditions, chronic illness, and disability are almost entirely absent from governance design, despite being primary subjects of clinical AI deployment. Frontline health workers — particularly those in under-resourced community health, rural, and Indigenous health settings — carry irreplaceable implementation knowledge that is not reflected in current governance frameworks. Indigenous Peoples and communities hold data sovereignty claims that exist prior to and independent of national AI governance frameworks. Their inclusion cannot be achieved through standard civil society consultation mechanisms; it requires recognition of governance authority, not just advisory participation. The Dialogue should include formal representation from Indigenous data governance bodies with the same standing as national delegations on provisions that affect Indigenous data. In the Global South broadly: national AI governance capacity is extremely uneven, and the loudest voices in international forums tend to be those with the resources to send delegations and produce policy papers. The Dialogue should actively commission governance input from practitioners in under-resourced settings and fund their participation. Among professional communities: social workers, community health workers, teachers, and public administrators who interact daily with AI-adjacent systems in benefit determination, triage, and assessment are almost entirely absent from AI governance discourse. Their knowledge of how algorithmic systems actually behave in practice — where they fail, what they miss, how practitioners compensate — is essential and currently invisible. Inclusion mechanisms: funded participation, asynchronous input options, translation into working languages beyond the six UN official languages, and formal co-design roles rather than consultation-only status.

The most persistent failure mode in international AI governance convenings is the production of documents rather than the development of shared understanding. Format innovation should address this directly. Deliberative polling — where participants engage with balanced, expert-informed briefings before expressing governance preferences — has a strong track record for complex technical-political questions. It surfaces genuine priority distributions rather than the positions of the most organized or best-resourced delegations. It would be particularly useful for cross-cutting questions where technical and ethical dimensions intersect. Red-team exercises, where mixed groups of technologists, ethicists, affected community members, and regulators work together to identify failure modes in proposed governance frameworks, generate more actionable output than panel discussion. They also build the kind of shared technical literacy across sectors that durable governance requires. For the health AI domain specifically: clinical simulation formats — where participants work through governance decisions using realistic patient scenarios — make abstract governance questions concrete and reveal implementation gaps that policy-level discussion obscures. This format works well across professional backgrounds and builds empathy between technical developers and clinical users. Structured adversarial formats — where one group defends a proposed governance standard and another challenges it from the perspective of affected communities — can surface tensions more honestly than consensus-seeking panels. For ongoing engagement between Dialogue sessions: structured written consultation with genuine response mechanisms (where submissions receive substantive replies, not acknowledgment receipts) maintains momentum and ensures the Dialogue accumulates knowledge rather than starting fresh each cycle. Critically, all format innovation must be accompanied by genuine language access. The most innovative engagement format fails if non-English-speaking participants are structurally disadvantaged in real-time deliberation. Investing in simultaneous interpretation and pre-circulated materials in multiple languages is not logistical overhead — it is governance design.

The most persistent failure mode in international AI governance convenings is the production of documents rather than the development of shared understanding. Format innovation should address this directly. Deliberative polling - where participants engage with balanced, expert-informed briefings before expressing governance preferences - has a strong track record for complex technical-political questions. It surfaces genuine priority distributions rather than the positions of the most organized or best-resourced delegations. It would be particularly useful for cross-cutting questions where technical and ethical dimensions intersect. Red-team exercises, where mixed groups of technologists, ethicists, affected community members, and regulators work together to identify failure modes in proposed governance frameworks, generate more actionable output than panel discussion. They also build the kind of shared technical literacy across sectors that durable governance requires. For the health AI domain specifically: clinical simulation formats - where participants work through governance decisions using realistic patient scenarios - make abstract governance questions concrete and reveal implementation gaps that policy-level discussion obscures. This format works well across professional backgrounds and builds empathy between technical developers and clinical users. Structured adversarial formats - where one group defends a proposed governance standard and another challenges it from the perspective of affected communities - can surface tensions more honestly than consensus-seeking panels. For ongoing engagement between Dialogue sessions: structured written consultation with genuine response mechanisms (where submissions receive substantive replies, not acknowledgment receipts) maintains momentum and ensures the Dialogue accumulates knowledge rather than starting fresh each cycle. Critically, all format innovation must be accompanied by genuine language access. The most innovative engagement format fails if non-English-speaking participants are structurally disadvantaged in real-time deliberation. Investing in simultaneous interpretation and pre-circulated materials in multiple languages is not logistical overhead - it is governance design.