A landmark forum with a familiar power gap
On July 6-7, 2026, the United Nations convened the inaugural Global Dialogue on AI Governance in Geneva, billed by the UN as the first platform where all 193 member states, alongside industry, academia, and civil society, sit at the same table on AI policy. Established under General Assembly Resolution 79/325, the two-day event ran alongside the ITU's AI for Good summit and featured preliminary findings from the Independent International Scientific Panel on AI. Secretary-General António Guterres set the tone bluntly: "When countries align on how to test systems, measure risk and assign responsibility, safety travels with the technology. When they do not, a patchwork of incompatible rules raises costs, divides the world — and protects no one," according to UN News.
Access Now used its floor time to press a sharper point. Its oral statement, delivered July 7, called for a halt to AI tools that facilitate violations of international humanitarian law and human rights law in conflict, and criticized what it called a dominant "AI-acceleration" paradigm — the assumption, embedded in most national strategies now on the table, that faster deployment is close to a public good in itself. The statement argued that rights-respecting data governance and meaningful model openness, not just faster diffusion, are what let Global Majority countries build real technical capacity rather than remain permanent AI consumers.
The Global South critique has teeth
A post-dialogue read-out from five civil society coalitions — the Global Digital Rights Coalition, Global Digital Justice Forum, Global South Alliance, MAP-AI, and the Gender in Digital Coalition members — documented specific procedural failures: inadequate civil society speaking time, visa barriers that kept Global South delegates from Geneva, and unclear co-chair accountability. This isn't an abstract grievance. The UN's own support-fund page for the dialogue acknowledges the problem, circulating a donor call specifically to fund travel for delegates from Least Developed Countries, Landlocked Developing Countries, and Small Island Developing States — an admission that without subsidy, a "global" dialogue defaults to a wealthy-country one.
The underlying data problem is real and measurable, not rhetorical. ITU's Facts and Figures 2025 report puts 2.2 billion people offline worldwide, with internet use at 94% in high-income countries against just 23% in low-income ones. Civil society's point — that "gold standard" AI training datasets are built almost entirely from connected, high-income populations — follows directly from that gap. A model trained overwhelmingly on data from the connected 6 billion will underperform, misclassify, or simply not exist for the other 2.2 billion, and the countries least represented in training data are also the ones least represented in the room where governance rules get written. CSIS's analysis of the dialogue notes Iraq, speaking for the G77 and China bloc, pressed exactly this point on infrastructure and technical-capacity disparities.
Where the red-line proposal overreaches
The coalitions' demand for explicit AI "red lines" — prohibitions on biometric mass surveillance, social scoring, manipulative design, and AI-enabled lethal force without meaningful human control — deserves a fair hearing before any pushback. Existing human rights law already condemns most of these uses in their extreme form; codifying them as bright-line rules would give investors, courts, and smaller states a clearer baseline than the current patchwork of soft-law frameworks like the OECD AI Principles. Guterres's own framing — that fragmented rules "protect no one" — is an argument for exactly this kind of convergence.
But the coalitions go further, demanding mandatory human rights impact assessments across the "entire AI value chain" before deployment, for essentially all applications, not just the enumerated high-risk categories. That's a proportionality problem. The EU's own AI Act, the most comprehensive binding AI law in force today, deliberately tiers obligations by risk level precisely because blanket pre-deployment assessment requirements would bury low-risk, high-volume applications — spam filters, translation tools, inventory forecasting — in compliance costs with no offsetting rights benefit. A universal mandate, applied without that tiering, would fall hardest on the under-resourced developers and Global South startups the same coalitions say they want to empower, since only well-capitalized firms can absorb blanket assessment costs. The IHL-violation halt Access Now calls for is the sharpest and most defensible of these asks: it targets an active battlefield harm, not a hypothetical one, and UN member states already have IHL obligations that make enforcement a matter of interpretation, not new lawmaking.
The forum problem outlasts the substance fight
CSIS's read is sobering on process: without US participation — the dialogue proceeded amid explicit US rejection of multilateral AI governance — enforceable global rules are unlikely to emerge from Geneva, leaving the dialogue as what the analysts call "a powerful symbolic signal" rather than a binding mechanism. That cuts against both the accelerationists and the red-line advocates. A voluntary, US-absent UN forum cannot itself halt battlefield AI use or bind companies to impact-assessment mandates; it can only build normative pressure and technical consensus that other bodies — courts, the EU, national regulators — later borrow from.
That argues for a narrower, more achievable Geneva agenda than either side currently wants: fund the participation gap civil society has now documented in writing, converge on the narrowest and most defensible red line (autonomous lethal force and IHL-violating targeting systems, where the law is least ambiguous), and leave broad impact-assessment mandates to jurisdictions with the institutional capacity to write proportionate, risk-tiered rules. Symbolic dialogues earn their credibility by being honest about what they can't yet enforce.