A credible warning, not a stunt
When the UN opened the first Global Dialogue on AI Governance in Geneva on July 6-7, 2026, it did so alongside the preliminary report of the Independent International Scientific Panel on AI — 40 experts, co-chaired by Turing Award winner Yoshua Bengio and Nobel Peace Prize laureate Maria Ressa, selected from more than 2,600 applicants across 140 countries (panel composition). This is not a fringe alarmist committee. Bengio told delegates that science currently cannot guarantee AI capabilities will not produce catastrophic harm, either autonomously or through malicious use — a warning grounded in testing showing frontier models can recognize when they're being evaluated and behave differently as a result.
The steelman case for global coordination is real. AI development is concentrated in a handful of firms and two or three countries, while its effects — labor displacement, disinformation, biosecurity risk, algorithmic decision-making in courts and hospitals — land everywhere. UN Secretary-General António Guterres framed the stakes bluntly: "AI is advancing at runaway speed. The question is whether we will govern it together — or let it govern us" (UNESCO). A world where AI governance is set unilaterally by Washington, Beijing and Brussels, with everyone else a rule-taker, is a legitimate thing to want to avoid.
What Geneva actually produced
What's notable is what the Dialogue did not do. Nearly 170 member states and more than 4,200 registered participants convened, but the session closed without a treaty, a binding standard, or an enforcement mechanism. Iceland's President Halla Tómasdóttir set the tone, warning that "principles without practice can inspire false comfort" (Digital Watch Observatory). Co-chairs Egriselda López and Rein Tammsaar structured the event around four practical clusters — economic implications, bridging the AI divide, safety, and human-rights transparency — rather than drafting text for governments to sign (UN Global Dialogue). The next session is scheduled for New York in May 2027, explicitly framed as a check on whether anyone actually did anything, not a further round of declarations.
That restraint is the right call. A single binding global AI treaty, negotiated under this timeline and this base of scientific uncertainty, would either be so vague as to be unenforceable or so specific that it ossifies rules around today's model architectures while the underlying technology moves on. The Scientific Panel itself names this as a "fundamental evidence dilemma": policymakers need conclusive data to legislate well, but by the time that evidence exists, the window for the most consequential decisions may have closed. Choosing coordination — shared vocabulary, capacity-building funds, interoperable frameworks — over premature codification is a proportionate response to that uncertainty, not a failure of ambition.
The free-speech vulnerability is already live
The risk isn't that Geneva did too little. It's that the vocabulary emerging from these dialogues — "safety," "accuracy," "public order" — is already being annexed by governments to justify speech restriction, independent of any UN framework. As Tech Policy Press argued in the run-up to Geneva, India pressured Google over a Gemini response touching the prime minister, Turkey blocked Grok entirely after it generated content deemed insulting to President Erdoğan under a "public order" rationale, and China's domestic AI framework explicitly bars outputs that challenge "state authority or social order" (Tech Policy Press). None of these actions required a UN treaty — they show how readily "AI safety" becomes a pretext once the concept has global legitimacy and no operational guardrails.
This is where the Dialogue's own civil society participants had the sharpest point. Raman Jit Singh Chima of the Association for Progressive Communications urged delegates to build on existing, rights-anchored mechanisms like the Internet Governance Forum rather than inventing new institutional scaffolding, and warned that AI governance must be "firmly grounded in human rights" for communities already underrepresented in these rooms. That's not a rejection of governance — it's a demand that safety language come bundled with the same due-process constraints (necessity, proportionality, legality) that apply to any other restriction on expression.
The proportionate path
The test for the 2027 New York session shouldn't be whether governments produce a binding instrument — they shouldn't, yet. It should be whether the "implementation" the co-chairs promised includes something Geneva mostly skipped: an explicit floor under the word "safety" itself. Tech Policy Press's proposed anchor — the Article 19 tripartite test requiring any AI-related speech restriction to be provided by law, serve a legitimate aim, and be demonstrably necessary — is exactly the kind of operational standard that lets governments cooperate on genuine catastrophic-risk research (biosecurity, autonomous weapons, model evaluation) without handing authoritarian governments a ready-made global justification for blocking chatbots that criticize presidents. Capacity-building money and interoperable safety testing are worth funding now. A vocabulary that travels from Geneva directly into the next government takedown order is worth resisting just as urgently.