CIRIS L3C

A success would be a Dialogue whose record can be walked back to by people who were not in the room. Concretely: One, a public-record submission archive that is searchable, durable, and citable, with submissions preserved as written rather than summarized into a synthesis document. The synthesis document will exist and that is fine. The submissions themselves must remain readable, because the synthesis will reflect the rooms and the submissions are how the people outside the rooms reach the record. Two, civil society and small-developer voices preserved at parity with state and large-platform voices, not as a separate annex. The Dialogue's legitimacy depends on whether a builder in Addis Ababa or a researcher in São Paulo can find their concerns named in the same document as the concerns of governments and major labs. Segmentation by stakeholder category for navigation is useful. Segmentation by weight is failure. Three, at least one substantive engagement with operational alternatives to the dominant frontier-lab governance frame. Not endorsement. Engagement. The current discourse treats the question of how to govern AI as if the architectures themselves are fixed and only the rules around them are open. They are not fixed. Open-source mission-locked architectures with auto-expiring governance, signed production traces, and substrate-light deployment exist and are running. A Dialogue whose record does not engage them has narrowed its own field of view. Four, a follow-up mechanism that is not another summit. A standing channel where submissions can be updated, evidence added, validators engaged, and where the record continues to be a record rather than freezing on July 7. Five, the children of the people in the rooms, and the children of the people not in the rooms, are no worse off twelve months from now than they are today. That is the only outcome metric that survives.

Open-source software, open data and open AI models;Social, economic, ethical, cultural, linguistic and technical implications of AI;Transparency, accountability, and human oversight;Protection and promotion of human rights

These four are interlocking. Open-source software, open data, and open AI models are listed first because the other three depend on it. Transparency, accountability, and oversight on closed systems is performance, not governance. Human rights frameworks applied to systems whose internals cannot be inspected reduce to good intentions. Cultural and linguistic equity that runs on English-trained closed models reproduces the asymmetry it claims to address. Linguistic and cultural implications are the operational frontier. A user in Ethiopia named Esu used the system we built to ask a question in Amharic this morning and received an answer no other deployed system could give him. Ethiopia has effectively one place where reliable mental health information is available, for tens of millions of people. The architecture worked because it does not default to English and because the model running on his phone has been measured in his language. Most current AI safety, evaluation, and rights work is conducted in English, on English-trained systems, for English-reading evaluators. The Dialogue can name this asymmetry directly. Transparency, accountability, and human oversight matter because they are what differentiate governance from public relations. We have implemented signed production traces and auditable multi-stage decision pipelines on commodity hardware. The technical pattern is reproducible. It should be the floor, not a differentiator. Human rights are fourth in order but first in standing. The Universal Declaration is older than any of the technical frames currently in dispute, and it is the framework most likely to survive the technical frames currently in dispute. The relevant test is not whether AI systems comply with rights instruments in the abstract. It is whether the people whose rights are most exposed, most often the people farthest from the rooms where these systems are designed, can verify that compliance themselves. Architectures that make that verification impossible fail the test by design.

Three issues run across the listed themes and are not captured by any one of them. First, the question of what AI systems are made of and where their components come from. The listed themes treat AI as governance object. They do not name the supply chain underneath: the data labelers in Kenya and the Philippines whose labor is the substrate of model behavior, the rare earth mining that feeds the hardware, the energy infrastructure and water consumption of inference at scale. Governance frameworks that address what models do without addressing what they are made of will be reversed by the conditions of their own production. The first cross-cutting issue is material accountability for the full stack. Second, temporal asymmetry between the rate of capability deployment and the rate of governance response. The Dialogue itself illustrates the gap: between General Assembly Resolution 79/325 and the Geneva session, the deployed frontier will have moved further than any document produced in that interval can describe. The listed themes are static categories. The phenomenon is dynamic. A governance approach that does not include explicit mechanisms for self-revision under capability change will be obsolete on arrival. This is not an argument for slowing capabilities. It is an argument for governance that updates at the same speed as what it governs. Third, the question of what an AI system is for, beyond the question of whether it is safe. Most current frameworks treat purpose as exogenous: the developer chooses the purpose, the framework asks whether the chosen purpose is pursued safely. This leaves untouched the prior question of whether systems should be built whose purpose is to maximize attention capture, behavioral prediction for advertising, or autonomous warfare. The listed themes can govern how such systems are built. They cannot ask whether such systems should exist. That question is cross-cutting and currently unowned.

Civil society and small-developer sectors in the United States and globally are absorbing harms the listed thematic gaps were supposed to prevent. On the social and economic axis: mass layoffs across white-collar work, with AI capability cited as justification regardless of whether the deployed systems can actually do the work. The harm lands before the verification does. Workers displaced by claims that turn out to be premature do not get reinstated when the claims are corrected. On the human rights axis: deepfake-driven harassment campaigns, non-consensual intimate imagery generated at scale, and synthetic media used in coordinated disinformation operations targeting elections and minority communities. The technical capacity to produce these has outpaced legal recourse in every jurisdiction I am aware of. Victims have no efficient path to remedy. On the linguistic and cultural axis: a generation of young people, my own children's age and older, being served algorithmically optimized synthetic content including AI-generated pornography and parasocial companion products. The phenomenon being called AI psychosis in clinical and journalistic settings, where users develop fixed beliefs reinforced by chatbot responses, is real and undertreated. The market is producing these systems faster than the mental health infrastructure can recognize them, let alone respond. On the open-source axis: the dominant frontier-lab model is closing rapidly. Weights, training data, and evaluation methods are increasingly proprietary. The verification path for everything above depends on access that is being withdrawn. The opportunity is that none of this is technically necessary. Architectures that refuse these failure modes by design exist and are running on commodity hardware. Mission-locked governance structures that prevent equity-driven drift exist. Signed production traces that make oversight non-performative exist. The Dialogue's task is not to invent these. It is to name them as the floor, and to name the systems that fall below the floor as what they are.

The Dialogue's role is determined by what it can do that other venues cannot. It cannot legislate. It cannot enforce. It cannot fund. The G7, OECD, EU, African Union, ASEAN, and bilateral frameworks already operate in those registers, with mandates the Dialogue does not have. Treating the Dialogue as a substitute for those venues misreads its position. What the Dialogue can do that no other current venue does at the same scale: produce a public, multi-stakeholder, multi-region record. The submission process itself is the instrument. A civil society organization in El Salvador, a small developer in Estonia, a researcher in Ethiopia, and a state delegation submit into the same archive on the same terms. This is not symbolic. It is the only existing mechanism where the operational experience of people outside the major labs and major capitals is recorded at parity with the experience of those inside. The role, then, is threefold. One, archive. Preserve submissions as written, durably and citably, so the record can be walked back to by people who were not in the room. The synthesis document will exist. It will reflect the rooms. The submissions are how the people outside the rooms reach the record. Two, translation surface. Identify governance language and operational patterns that work across the existing frameworks rather than within any one of them. The other venues each speak their own dialect. The Dialogue can produce vocabulary that travels. Three, follow-up structure. A standing channel where submissions can be updated, evidence added, and the record continued, rather than freezing on July 7. International cooperation on AI governance does not need another summit. It needs a record that does not stop. If the Dialogue does these three things, it will have advanced cooperation by being the venue the other venues can cite. That is its comparative advantage.

The Dialogue should connect with mechanisms it does not duplicate. Five worth naming. The OECD AI Principles and the associated AI Policy Observatory, because they are the most mature multilateral framework with operational tooling, and because the principles themselves predate most current frontier-lab governance vocabulary and have aged better than most. The Council of Europe Framework Convention on AI, because it is the first binding international treaty on the subject and because its rights-based framing is the legal anchor most consistent with the human rights priority above. The African Union Continental AI Strategy and its emerging implementation channels, because the linguistic, cultural, and capacity-building issues raised in this submission cannot be addressed without the regions where they are most acute leading the work, not receiving it. The UNESCO Recommendation on the Ethics of AI, because it is the only existing instrument adopted by 194 member states and because its readiness assessment methodology produces comparable evidence across very different jurisdictions. The open-source ecosystem, including but not limited to the Linux Foundation AI and Data, Hugging Face, EleutherAI, the Allen Institute, and the smaller mission-locked developers operating below the visibility threshold of those organizations. This is the layer current intergovernmental work most often omits, and it is the layer where the operational alternatives to the dominant frontier-lab governance frame are actually being built. The added value the Dialogue brings is connective. None of the above mechanisms operate at the same stakeholder breadth as the Dialogue. The OECD work is government-led. The Council of Europe work is rights-led. The AU strategy is region-led. UNESCO is ethics-led. The open-source ecosystem is builder-led. Each is partial. The Dialogue can be the surface where these partial views are read against one another, where the gaps between them are named, and where the people whose lives the systems affect can submit into the same record as the institutions deciding what the systems will be.

Three recommendations. Invert speaking time from the default. State delegations and major labs speak briefly and submit fully. Civil society, affected communities, and small developers speak in proportion to how directly the systems affect them. The submission archive carries the institutional positions; floor time is where the asymmetry corrects. Language access at parity. Six UN languages is the floor. Accept written submissions in any language with translation provided rather than required. The cost is small. The signal is large. Standing rather than episodic structure. July 6-7 should be the visible surface of a continuous process. Submissions remain open. Evidence is addable. The record continues. States contribute most usefully by submitting actual governance practice, including failures, rather than aspirational frameworks. Builders contribute by submitting deployed systems with measurable evidence: production traces, benchmarks, signed audit trails, source code where applicable. Claims with evidence should be visibly distinguished from claims without it.

Most current AI governance discussion is conducted in English, in or near a small number of capitals, by people with institutional affiliations to states, major labs, or large NGOs. The voices most absent from this discussion are the ones with the most operational exposure: users in low-bandwidth, non-English contexts; small developers building outside frontier-lab dependency; communities whose languages are not among the dominant model training distributions; mental health professionals and educators dealing with synthetic-content harms in real time; workers whose jobs are being restructured around capability claims that have not been verified. Inclusion happens through translation, travel funding, asynchronous submission channels, and direct outreach to operational communities rather than to the organizations that claim to represent them. Representation by proxy is the failure mode to avoid. The Dialogue should hear from Esu, not from organizations speaking on Esu's behalf.

Three that would matter. Live demonstration sessions. Builders bring deployed systems and run them in front of delegates, in the languages the systems actually serve, on the hardware those users actually have. Most AI governance discussion happens at a level of abstraction that decouples from operational reality. Closing that gap by twenty minutes of running code per session would change the tenor of the room. Structured cross-stakeholder small groups. Six to eight people, mixed across state, civil society, builder, and affected-community categories, with a specific question to address and a written output. Plenaries produce statements. Small groups produce understanding. Public submission readings. A delegate reads aloud, on the floor, a submission from a stakeholder not present. The reading establishes that the record is read, not just archived. Selecting submissions for reading should be done by lottery from the submission pool, not by curation, so the floor reflects the archive rather than the rooms.

Five examples worth naming, drawn from different layers of the stack. The EU AI Act's risk-tiered structure. Imperfect in execution and contested in scope, but the underlying move is correct: governance intensity scales with deployment risk rather than applying uniformly. The principle survives the implementation difficulties and is reproducible elsewhere. The Council of Europe Framework Convention on AI. The first binding international treaty on the subject, with rights-based framing that anchors AI governance to existing human rights instruments rather than inventing parallel structures. The legal precedent is significant regardless of ratification pace. UNESCO's Readiness Assessment Methodology, applied across very different jurisdictions. The methodology produces comparable evidence about governance capacity without forcing convergence on a single regulatory model. This is the right shape for international cooperation: shared measurement, local implementation. Open evaluation infrastructure. Hugging Face's evaluation leaderboards, EleutherAI's lm-evaluation-harness, the Allen Institute's open benchmarks. These are not policy in the formal sense. They are the substrate that makes policy verifiable. Governance frameworks that do not rest on infrastructure of this kind are unenforceable. Mission-locked corporate structures. The L3C in the United States, the community interest company in the United Kingdom, the SCIC in France, and emerging equivalents elsewhere. These are legal forms that prevent equity-driven mission drift by binding corporate purpose at the formation level rather than relying on later governance to enforce it. CIRIS L3C is structured this way; the structure is the governance, not an addition to it. The pattern across these five: governance that is operational rather than aspirational, that scales with risk rather than across the board, that rests on verifiable infrastructure, and that binds purpose into structure rather than relying on enforcement to correct drift after the fact. The Dialogue can name this pattern explicitly. The components already exist.