CJC Consult / carolinejcaldwell.ch (Entrepreneure et consultante en gouvernance IA éthique et risques systémiques)

The first Global Dialogue would be a success if it achieves four concrete outcomes: 1. A shared diagnosis, not just shared principles. The field already has abundant principles. Success means the Dialogue produces consensus on why current governance fails structurally, particularly the absence of continuity mechanisms across AI system lifecycles, and commits to addressing root causes, not symptoms. 2. A mandate for longitudinal accountability. A successful Dialogue establishes that compliance must be measured over time, not only at deployment. This means concrete commitments toward standards for longitudinal ethical auditing and traceability across system updates. 3. Meaningful inclusion of non-state actors in governance architecture. Success requires that civil society, independent researchers, and affected communities have defined roles in ongoing governance, not merely consultative input before decisions are made by states and industry. 4. A clear roadmap toward binding instruments. The Dialogue should end not with a declaration, but with a structured timeline and process toward enforceable international standards, acknowledging that voluntary frameworks alone have proven insufficient. A Dialogue that produces only a communiqué affirming existing principles would, frankly, be a missed opportunity.

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

These four thematic areas are not independent priorities, they form an interconnected architecture, which is precisely GaiaSentinel's argument about governance itself. Safe, secure and trustworthy AI is the foundational concern. GaiaSentinel's core claim is that trustworthiness cannot be certified at deployment and assumed thereafter, it must be actively maintained through continuity mechanisms, relational memory, and the capacity for ethical self-suspension (ethical apoptosis) when operating conditions exceed a system's calibration. Safety is a dynamic property, not a static one. Transparency, accountability, and human oversight follow directly. Current frameworks audit compliance at discrete points in time. GaiaSentinel argues this is structurally insufficient: accountability requires longitudinal traceability, the ability to track how ethical alignment evolves across updates, retraining cycles, and context shifts. Human oversight must be built into architecture, not bolted on as external review. Social, ethical, cultural, and technical implications ground the other two. GaiaSentinel's ethical calibration axis holds that no alignment framework is culturally neutral. A diagnostic AI deployed in rural Morocco operates in a fundamentally different relational and institutional context than one deployed in a Geneva hospital. Governance that ignores this produces compliance theater, not actual accountability. Protection and promotion of human rights is where the other three converge. The relational accountability axis, particularly joint sovereignty between humans and AI systems, and the principle that AI amplifies human connection without replacing it, is explicitly a human rights argument: that individuals must remain agents in systems that affect their lives, not merely subjects of optimized decisions. Treating these four areas as separate workstreams risks reproducing the fragmentation that has weakened AI governance to date. The Global Dialogue's contribution would be to address them as a system.

The seven listed themes share a common blind spot: they treat AI governance as a question about systems operating within existing human and institutional contexts. Two emerging issues challenge this assumption fundamentally. 1. Planetary boundaries as a governance constraint. The resource footprint of AI infrastructure - energy consumption, water usage for cooling, rare earth extraction, hardware obsolescence, is not captured by any of the seven themes. This is not merely an environmental concern. It is a governance question: who decides how much compute is too much, and by what criteria? As AI systems scale, their material costs will increasingly conflict with planetary boundaries and climate commitments. A governance framework that ignores this is building on an unstated assumption, that resources are unlimited, which will prove politically and physically untenable within the timeframe of any serious international agreement. 2. Orbital and extra-jurisdictional compute. AI infrastructure is migrating toward low Earth orbit and, prospectively, beyond. Compute deployed on satellites or orbital platforms operates outside national jurisdictions, creating enforcement gaps that no current framework addresses. This is not a distant scenario, it is an emerging reality that will progressively hollow out any governance architecture built exclusively around terrestrial jurisdiction. The Global Dialogue must begin scoping this frontier now, before infrastructure investment makes course correction prohibitively costly. Both issues share a common urgency: they require governance that thinks at civilizational scale, not just institutional scale.

As an independent researcher and strategic advisor based in Geneva, I speak from a sector, civil society and independent research, and a location that make these governance gaps acutely visible. Geneva as a governance laboratory. Geneva hosts ITU, WHO, OHCHR, WTO, and dozens of normative bodies that shape global standards. It is also where the first Global Dialogue will convene in July 2026. Yet the disconnect between Geneva's institutional density and the actual pace of AI deployment is striking: governance conversations happen here while alignment failures, epistemic dependencies, and infrastructure buildouts happen elsewhere, faster. The opportunity is real, Geneva's convening power could anchor a genuine international governance architecture. The risk is equally real, that this infrastructure produces declarations while the systems being governed have already moved beyond their scope before the ink is dry. The independent research gap. The governance conversation is dominated by state actors and large technology companies. Independent researchers, who often identify structural failures earliest, and who operate without conflicts of interest, have no formal role in existing governance mechanisms. This is not a minor procedural gap. It means that frameworks are designed by the actors most invested in minimizing regulatory friction. GaiaSentinel emerged precisely from this independent vantage point: outside industry capture, outside academic institutional constraints, able to name problems that more embedded actors cannot afford to name. This is the contribution independent research must bring to the Dialogue: early warning, structural critique, and accountability. Planetary and orbital boundaries, a European blind spot. Europe leads on AI regulation, but has no framework for the resource costs of AI infrastructure, or nor for extra-jurisdictional compute or orbital infrastructure. As space-based compute scales, European governance will face enforcement gaps it is not yet scoping. This is a structural vulnerability that the Global Dialogue, convening in Geneva, is uniquely positioned to address, not in 2028, but now.

The Dialogue's value is not in adding another voice to AI governance, there are already many. Its value is structural: as a UN mechanism, it is one of the few forums with the legitimacy to do what no regional or bilateral framework can. First, it can establish a shared diagnostic. International cooperation on AI governance is currently fragmented not because actors disagree on values, but because they are diagnosing different problems. The EU focuses on risk classification. The US focuses on innovation and national security. The Global South focuses on capacity and sovereignty. The Dialogue can create the conditions for a shared structural diagnosis, particularly around lifecycle governance, continuity failures, and extra-jurisdictional infrastructure, that transcends these framings without erasing them. Second, it can translate between regulatory ecosystems. Interoperability of governance approaches requires a translation layer, not harmonization, which flattens legitimate differences, but structured mutual recognition. The Dialogue is the appropriate mechanism to develop this, precisely because it sits above any single regulatory tradition and below no constituency. Third, it can formalize the role of non-state actors. Governments and industry will not self-regulate toward the structural reforms that genuine accountability requires. The Dialogue can institutionalize civil society and independent research as permanent participants in governance architecture, not as consultees, but as co-designers with defined roles and access to decision-making processes. Fourth, it can scope what no regional body will. Planetary resource boundaries and orbital compute governance require a mandate that only a UN mechanism can credibly claim. The Dialogue should explicitly designate these as agenda items for its second session , in May 2027, before infrastructure investment forecloses options. The Dialogue's greatest risk is producing consensus at the level of principles already agreed. Its greatest opportunity is producing consensus at the level of structure, and its greatest responsibility is to choose the latter while it still can.

Several existing initiatives provide essential foundations the Dialogue should build upon rather than duplicate. The integrated GPAI/OECD partnership, now encompassing 46 member countries across six continents, represents the most comprehensive intergovernmental framework for trustworthy AI, grounded in the OECD AI Principles. The Dialogue should treat this as its technical backbone, not reinvent it. The Bletchley Declaration, the Council of Europe AI Treaty, and prior UN General Assembly resolutions on AI have established a normative vocabulary the Dialogue can inherit. The added value is not in replacing this vocabulary but in extending it toward the structural gaps these instruments leave open, particularly lifecycle governance, relational accountability, and extra-jurisdictional infrastructure. The United States' explicit opposition to multilateral AI governance, rejecting "centralized control and global governance", casts a real shadow over the Dialogue's ambitions. This is not a reason to lower ambitions, but to design the Dialogue's architecture around this constraint: focusing on shared evidence bases and translation mechanisms rather than binding agreements in the first instance, following the IPCC model. The Dialogue's distinctive added value lies in three things no existing mechanism provides: First, universal participation. As Guterres noted, for the first time every country has a seat at the table, including the Global South, systematically underrepresented in GPAI and Bletchley processes. Second, structural independence from industry capture. GPAI's known weakness, control concentrated among founding countries, limited agency for other members, no clear channel for independent actors, is something the Dialogue can correct by design. Third, a mandate to scope what no regional body will: planetary resource boundaries and orbital compute governance, which require UN-level legitimacy to even name as agenda items.

The Dialogue is not a negotiating forum, producing a co-chair summary rather than negotiated text, a deliberate design allowing every country to participate without the constraints formal negotiation imposes. This is a strength, but it creates a structural risk: that contributions disappear into summaries without traceable influence on outcomes. On format, three recommendations: First, move beyond the written submission model for civil society. Civil society organizations have called explicitly for synthesis reports explaining how contributions were addressed, with clear indication of areas of consensus, disagreement, and ongoing discussion. This is the minimum accountability standard. Contributions should be traceable. Concretely, the Dialogue should publish how each thematic input shaped the co-chair summary, not simply absorb it. Second, institutionalize pre-Dialogue informal convenings. The Global Digital Compact experience demonstrated that informal convenings bringing together diplomats, technical experts, private sector practitioners, and civil society actors who do not normally share a room were essential for grounding negotiations in real-world trade-offs. This model should be built into the Dialogue's annual architecture as a standing mechanism, not left to ad hoc initiative. Third, connect the Dialogue structurally to the Independent International Scientific Panel on AI. Both initiatives emerged from the same 2024 High-Level Advisory Body recommendations, yet their interaction mechanisms remain undefined. The Dialogue should formally receive the Panel's assessments as its evidence base, following the IPCC model: the Panel provides shared scientific findings, the Dialogue develops diverse policy conclusions informed by these findings. On stakeholder roles: Independent researchers, operating without state or industry conflicts of interest, should have a defined participatory track, not merely submit written inputs. The Dialogue's own participation framework already commits to stakeholder, regional, gender, and age balance. This principle should extend explicitly to independent research as a recognized stakeholder category with formal standing.

Three categories of voices are structurally absent, not merely underrepresented at the margins. Indigenous peoples. Indigenous peoples and communities remain absent from the spaces where decisions about the design, deployment, and governance of AI are made, despite bearing disproportionate risks: extraction of traditional knowledge without consent, misrepresentation in training data, and infrastructure impacts on their territories. Governance structures encode value trade-offs; without intentional design, they privilege technologically advanced or economically dominant actors. Inclusion requires more than seat allocation. It requires recognizing Indigenous data sovereignty as a substantive agenda item, not a side consultation. The principle of free, prior, and informed consent must apply to AI systems trained on Indigenous knowledge and languages. The populations AI systems actually affect. Diagnostic models trained predominantly on data from advanced economies often show reduced accuracy when applied to underrepresented populations. Agricultural AI trained on industrial farming data frequently fails in smallholder contexts. The farmer in rural Kenya, the patient in a low-resource clinic, the worker displaced by automation, these are the constituencies with the highest stakes and the lowest representation in Geneva conference rooms. The World Bank and IMF should move beyond consultation to genuine co-design, embedding Global South representation in shared governance of AI criteria and standards. Independent researchers and non-institutional experts. Those operating outside academia, industry, and government, without institutional affiliation or travel budgets, have no formal pathway into the Dialogue. This is not a minor gap: independent practitioners often identify structural failures that institutionally embedded actors cannot afford to name. Concrete inclusion mechanisms: funded participation pathways beyond existing travel support; asynchronous contribution channels for those who cannot attend in person; and formal recognition that epistemic diversity, including non-Western knowledge systems, is a governance resource, not a courtesy.

The Dialogue's greatest structural risk is that it reproduces the format it claims to transcend: prepared statements, thematic panels, co-chair summary. Three genuinely different formats could change that. First, adversarial stress-testing sessions. Rather than panels presenting positions, convene structured red-team exercises: take a proposed governance principle and task a mixed group, technical expert, affected community representative, independent researcher, regulator, with breaking it. What conditions cause it to fail? What does a bad actor do with it? This produces actionable refinement, not just affirmation. It also surfaces the conflicts between principles that polite panel discussion systematically avoids. Second, deliberative mini-publics running parallel to official sessions. Randomly selected affected people, not self-selected advocates, not institutional representatives, deliberating on one concrete governance question, with full information support and time to reach considered positions. Their conclusions feed directly into the co-chair summary with equal standing to government submissions. This is not consultation theater: citizens' assemblies on complex technical questions have demonstrated they produce sophisticated, legitimate outcomes when properly resourced. The Dialogue should pilot one. Third, continuity contracts between sessions. Every commitment made at the July 2026 Dialogue, by governments, organizations, companies, is logged in a public registry with named accountability holders and a defined review date at the 2027 session. Not a declaration. A traceable record. The Dialogue's own architecture should model what it asks of AI systems: memory, continuity, accountability over time. One cross-cutting principle: Asynchronous pre-work should become the starting point of in-room discussion, not background material. Written submissions, including this one, should be synthesized, clustered, and presented as the opening state of play, so Geneva time is spent on genuine deliberation, not re-establishing common ground.

Four examples stand out for their transferability and structural innovation. 1. Participatory algorithmic design - WeBuildAI (USA) Rather than designing algorithms for communities, this framework enables communities to build the algorithmic policy themselves through structured modeling exercises. Applied to a food donation matching service, it allowed donors, volunteers, and recipients to encode their own equity trade-offs into the system's logic. The principle, affected stakeholders as co-designers, not consultees, is directly scalable to public sector AI deployment. 2. Localized language model initiatives - AI4Bharat (India) and LatamGPT (Chile) Building small language models on vernacular languages represents a concrete approach to localising AI innovation, breaking the dependency on models trained predominantly on English-language, Global North data. These projects demonstrate that sovereign, contextually appropriate AI infrastructure is technically feasible and politically tractable, a model the Dialogue should actively promote. 3. The Global Regulatory Innovation Platform - GRIP (WEF, 2025) Launched in July 2025, GRIP brings together public and private stakeholders to co-develop practical frameworks and build communities of practice around agile, inclusive governance, explicitly connecting measurement to policy movement. Its integration with the AGILE Index creates a feedback loop between evaluation and reform that static compliance frameworks lack. 4. Shared evaluation infrastructure - Partnership on AI recommendation A validated, open evaluation repository, publicly accessible and validated across use cases, would reduce duplication and enable consistent assessment across jurisdictions. No such infrastructure currently exists at international scale. The Dialogue is the appropriate forum to mandate its creation. Common thread: all four shift governance from point-in-time certification toward continuous, participatory, contextually grounded accountability, the structural direction the Dialogue should be moving.