Erisian Group

Success is not the production of a consensus text. AI governance is currently characterised by several incommensurable framings; sovereignty-first, rights-based, multilateralist, market-led, civilisational-risk, development-justice, and indigenous epistemological positions among them, held by serious institutions with serious constituencies. The intelligence the Dialogue needs actually lives in the fault lines between these divergent worldviews, not in the language they can all sign onto. Every prior multilateral AI process has aggregated to the mean or to power: procedurally, because the default reflex is to draft toward consensus often before divergence is mapped; structurally, because the tools to hold and surface the real scale of divergence have not existed. The result is instruments that signatories operationalise in incompatible ways, and a Dialogue whose own analytical inputs systematically miss what matters most. The first Dialogue will be a success if it achieves three things. First, it surfaces the fault lines and divergent worldviews. Where Member States and stakeholders actually diverge across the four thematic clusters, where shared language conceals incompatible assumptions, and which framings are being absorbed into others by default rather than by deliberation. This becomes the empirical baseline for everything that follows. Second, it shifts capacity-building from skills-and-infrastructure transfer to epistemic capacity. Every Member State, regardless of technological position, must have the analytical infrastructure to independently assess the global discourse, identify where its interests and worldviews are represented or omitted, and engage on equal footing in standard-setting. Open data, open models, and open-source software are necessary but not sufficient; open analytical capacity is the missing element. Third, it structurally protects the role of independent, non-captured analytical infrastructure in the governance process. The credibility of multilateral AI governance requires that the lens through which it sees itself is not owned by the parties it governs. If the Geneva session produces these three outcomes, the May 2027 New York session inherits a substrate strong enough to hold a real conversation. Without them, the Dialogue becomes another venue where the loudest or most powerful framings win by default and less-resourced positions are absorbed into language they cannot operationalise at home.

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

Interoperability of governance approaches is the area where the gap between rhetorical alignment and operational divergence is largest, and where the cost of that gap compounds fastest. Interoperability cannot be designed without a current, multi-jurisdictional map of where governance instruments actually differ, including where shared language conceals incompatible assumptions. The fault lines are the design substrate. Erisian's discourse-mapping infrastructure is built for precisely this gap. Interoperability that resolves these incompatibilities by collapsing them into lowest-common-denominator language is the suppression of governance pluralism in the name of harmonisation. Social, economic, ethical, cultural, linguistic and technical implications cannot be assessed in the aggregate. They are differently distributed by language, region, sector, and worldview, and the framings used to describe them shape which implications become visible to policymakers. Without a methodology that preserves framing diversity, the Dialogue's analytical inputs systematically privilege framings already best-resourced in English-language Anglo-American discourse. AI capacity-building as currently framed; the transfer of skills, infrastructure, and tools from advanced to developing states, risks deepening dependency on the data infrastructure and value frameworks of providing states. Capacity-building succeeds to the extent that it is paired with epistemic capacity: the analytical infrastructure that allows every Member State to read the global landscape independently and see the fault lines for itself. Erisian commits to making a stripped-down version of its discourse-mapping capability available to Permanent Missions and developing-country institutions on a non-commercial basis. Transparency, accountability, and human oversight cannot be imposed on AI systems if they are not first practised in the governance discourse about those systems. Making the discourse itself transparent by surfacing who is speaking, from which framing, with what assumptions, and where positions diverge- is the precondition for making the systems it governs transparent, accountable, and subject to meaningful oversight.

The cross-cutting issue not captured by the four clusters is what AI systems do to collective reasoning itself. AI systems collapse and synthesize plural human inputs into single outputs, with the collapse defaulting toward the statistical mean of the training data or toward the framings of those who shape alignment, system prompts, and deployment contexts. The disagreements that human governance, science, and law depend on as productive substrate are silently resolved before reaching the human evaluator. This is a structural property of current systems, not a bug to be patched. The implication for the Dialogue is direct. Safety frameworks cannot detect what they do not measure; the collapse of plural reasoning is a safety property currently absent from major evaluation suites. Interoperability that harmonises governance approaches by erasing their distinct commitments produces convergence, not interoperability. Capacity-building that transfers tools without preserving the reasoning traditions of the receiving context deepens monoculture. Open-source releases do not solve the problem if post-training alignment is monocultural. Human rights protections require visible disagreement; meaningful oversight is impossible when the disagreements have already been hidden inside the model's output. The Dialogue is itself an instrument for holding plural perspectives in productive tension. Its premise is that legitimate AI governance requires what AI by default destroys. Naming this - call it the preservation of dialectical capacity as a governance objective, gives the four clusters a frame they can be read through and a criterion against which the Dialogue's own outputs can be measured.

Erisian operates across the global foresight and policy sector and the governance gaps in the four selected areas are reshaping the sector in ways that compound each other. On interoperability, institutions across regions are being asked to align with frameworks whose surface language converges while their underlying assumptions diverge. Foresight and policy organisations in developing countries are absorbing this language without the analytical infrastructure to see where the assumptions inside it conflict with their own contexts. The result is institutional commitments that cannot be operationalised locally, and a slow erosion of distinct regional governance traditions into vocabulary they did not author. On capacity-building, the sector is seeing significant investment in skills training, compute access, and infrastructure transfer and almost none in the epistemic capacity to read the global discourse independently. Developing-country institutions are receiving tools without receiving the analytical autonomy to assess what those tools embed. Dependency is deepening under the language of empowerment. On transparency and oversight, the foresight and policy sector itself is increasingly opaque about which framings shape which outputs, which institutions are converging on which positions, and where the fault lines actually run. Sector actors are being asked to advise governments on AI transparency while the discourse they draw from is itself opaque. The compounding effect is that institutions outside the dominant discourse are progressively absorbed into framings they did not shape, equipped with tools that embed those framings, and asked to govern in language that obscures the divergences that matter most to their constituencies.

The Dialogue is the only venue with the universality and legitimacy to do what no regional or plurilateral initiative can: hold the global AI governance conversation in one room. What it does with that position is the open question. Three possibilities seem most consequential. It could serve as the empirical baseline for global AI governance by surfacing where Member State and stakeholder positions actually diverge, and where shared language conceals incompatible assumptions. Every regional and plurilateral initiative currently operates without knowing where its framings sit relative to others. It could shift capacity-building from tool transfer to epistemic capacity, pairing compute, skills, and infrastructure with the analytical autonomy for every Member State to read the global landscape independently. Without this, capacity-building deepens dependency under the language of empowerment. It could structurally protect the role of independent, non-captured analytical infrastructure in the governance process. The legitimacy of multilateral AI governance depends on the lens through which it sees itself not being owned by the parties it governs.