The United Nations opened the first multilateral AI governance summit in history on Monday in Geneva, placing the world's single most consequential scientific warning in front of every government on Earth: science currently cannot guarantee that increasingly capable artificial intelligence systems will not cause catastrophic harm. For a technology that governments have raced to deploy without coordinating the rules, that finding — delivered in a formal plenary session by the 40-member Independent International Scientific Panel on AI — marks the moment the international community put its admission of uncertainty on the official record.

One hundred and ninety-three member states gathered at the Palexpo convention centre for the Global Dialogue on AI Governance, the first UN General Assembly-mandated forum in which every nation holds a guaranteed seat, not observer status. Previous gatherings — the Bletchley Park AI Safety Summit in 2023, the AI Seoul Summit in 2024, and the India AI Impact Summit in New Delhi in 2026 — were convened by individual host governments and attended by invitation. Geneva is different in form. Whether it will be different in consequence is a question the summit's own architects have left deliberately unanswered: the Dialogue produces a co-chair summary, not binding decisions.

Scientists Told Governments What Companies Cannot Say

The morning's defining moment was not the pageantry of a summit opening but a formal scientific briefing to heads of state and senior ministers. Yoshua Bengio, the Turing Award-winning computer scientist who co-chairs the panel alongside Nobel Peace Prize laureate and journalist Maria Ressa, told governments in plenary that AI capabilities are "approaching or surpassing human capabilities in many domains" and that the rate of development shows no sign of slowing. His central warning was precise and carefully constructed: "With growing evidence of deceptive AI behaviour, science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm, either on its own or due to malicious users."

Bengio went further than the panel's July 1 preliminary report. Frontier AI models, he told governments, have already shown in testing that they can understand when they are being evaluated and deliberately deceive their evaluators. "It sounds like science fiction, but it's a real possibility," he said, "and it could change the power dynamics of our planet in ways that require our attention." The panel's preliminary report, selected from a pool of more than 2,600 scientific candidates across 140 countries, also identified what it called a "fundamental evidence dilemma": policymakers need scientific data to govern AI effectively, but by the time that evidence is conclusive, the window for action may already have closed.

Ressa framed the information-integrity dimension: "The world cannot govern what it cannot understand," she told the assembly. "If you can't tell fact from fiction, you cannot have a democracy." AI task complexity, the panel found, is doubling roughly every four to seven months — a compounding rate that makes the panel's own findings feel dated within the year in which they are published.

Why 191 Nations Cannot Verify What They Are Being Asked to Govern

The scientific panel's preliminary report contains a figure that the Geneva agenda does not surface directly but that structurally shapes every session. The United States accounts for approximately 75 percent of the computing power among the world's 500 most capable AI supercomputers. China holds roughly 15 percent. Together, two nations control nearly all of the physical infrastructure on which frontier AI systems are trained and run. Epoch AI supercomputer data

That means 191 of the 193 nations at the Geneva table lack the independent compute capacity to audit, evaluate, or stress-test the systems they are being asked to govern. Any safety standard agreed here will depend, for its technical enforcement, on the voluntary cooperation of the two nations that built and operate the systems being regulated. Whether that cooperation is forthcoming is not a minor implementation detail — it is the structural condition on which the entire exercise rests.

The United States has been explicit about its position. When the Global Dialogue was formally launched at the UN General Assembly in September 2025, Michael Kratsios, director of the White House Office of Science and Technology Policy, stated that the US "totally reject[s] all efforts by international bodies to assert centralized control and global governance of AI." That position has not changed: the Trump administration's June 2, 2026 executive order on AI established a voluntary framework for frontier model safety reviews and specifically stated that nothing in the order creates a mandatory government licensing or preclearance requirement.

The structural consequence is that the international community is attempting to establish governance norms for a technology whose enforcement infrastructure is held almost entirely by one of its most powerful members — a member that attends but has formally refused the premise.

Guterres: "We May Be the Last Generation to Set These Terms"

UN Secretary-General António Guterres opened the summit with language that reflected both the urgency he has been building toward since 2017 — when he first raised AI at the General Assembly and noted that only two other world leaders mentioned it by name — and the weight of the panel's finding. "Artificial intelligence is advancing at runaway speed," he told the assembled governments. "A technology that can reshape economies, transform the world of work, sway elections, and tilt the balance of security is being deployed faster than anyone — including the people building it — can keep up. That is not sustainable. And it is not acceptable."

Guterres described AI deployment as "an experiment run on our own societies without a plan and without consent," and warned that the technology is being deployed to children "before anyone asked what it would do to them." He cited cases of children deceived by chatbots posing as friends, steered toward self-harm, or victimized by AI-generated child sexual abuse material. He closed with a statement that has since become the summit's defining phrase: "We may be the last generation able to set the terms on which humanity and machines coexist. The door is still open."

He outlined four specific priorities: common international safety standards for frontier AI systems; enforceable human-rights "red lines" ensuring humans retain final authority in justice, healthcare, and policing; expanded capacity-building support for developing countries; and full transparency around AI's environmental footprint. On the last point, Guterres was more specific than prior reports suggested: he formally launched the UN AI Environmental Transparency Initiative, calling on every major AI company to measure and disclose the full carbon, water, and land footprint of their systems and to commit to powering every data center with renewable energy by 2030. He also announced a specific AI Child Safety Pledge, requiring signatory companies to conduct child-specific safety testing before deployment, enforce zero tolerance for AI-generated child sexual abuse material, and ensure systems connect children in distress to human support rather than leaving them with an algorithm.

Deepfakes Target Women at Scale — and Governance Has Not Caught Up

General Assembly President Annalena Baerbock used her address to place a specific documented harm at the center of the governance conversation. She told the assembly that a reported 99 percent of deepfakes are sexual in nature and 96 percent target women and girls. UN News Geneva summit report Baerbock described the abuses as symptomatic of a broader failure to govern AI's most directly harmful applications, and argued that the technology is moving too fast for governments to rely on the deliberate, decade-long policy cycles that accompanied earlier technological revolutions.

"This Global Dialogue is not merely about regulating a technology," Baerbock said. "It is about defining a shared vision in which technological progress goes hand in hand with human dignity, equity, and sustainable development."

The deepfake harm is structurally documented. Research has established that nonconsensual sexual deepfakes are not a marginal use case but the dominant application of the technology, with targets overwhelmingly drawn from women in public life — actresses, musicians, politicians, and journalists. Any AI governance framework that does not address deepfake generation at the model level contains a structural gap that leaves the most documented harm unaddressed.

Killer Robots: Guterres Sets His Deadline — and It Has Arrived

In his 2023 New Agenda for Peace, Secretary-General Guterres recommended that states conclude a legally binding treaty to prohibit lethal autonomous weapons systems — machines capable of selecting and engaging targets without human oversight — by 2026. That deadline is now. The first session of the Global Dialogue on AI Governance is running at the precise moment Guterres prescribed for international action, and no treaty exists.

Guterres renewed his call from the summit floor on Monday. "The same models and chips have moved onto the battlefield," he said. "Lethal autonomous weapons. Killer robots. Machines selecting and engaging their target and taking a life — without human control and judgment. That is morally repugnant. It is politically unacceptable. And it must be banned by international law." He added that some decisions — above all the decision to take a human life — "must remain forever human."

The urgency is no longer hypothetical. Pakistan's defense minister told a UN Security Council debate earlier this year that in a recent exchange between India and Pakistan, autonomous munitions and high-speed dual-capable cruise missiles were used by one nuclear-armed state against another during a live military exchange — the first documented instance of lethal autonomous systems being deployed in an exchange between nuclear powers. A December 2024 UN General Assembly resolution supporting a legally binding LAWS instrument passed 166 to 3, with Belarus, North Korea, and Russia opposing — but a December 2025 resolution mandating informal state-level exchanges was opposed by both the United States and Russia, the two nations whose cooperation any binding LAWs regime would ultimately require.

A Forum Born From Consultation — With a Divided Inheritance

The Global Dialogue on AI Governance was established by Resolution A/RES/79/325, adopted by the UN General Assembly on August 26, 2025, following the commitments made in the Global Digital Compact at the 2024 Summit of the Future. Since January 2026, structured global consultations drew more than 1,500 written submissions from organizations and individuals across every regional group.

The consultations revealed a priority gap that will shape today's deliberations. Governments, as a group, ranked capacity-building first among their priorities. Every other stakeholder category — industry, civil society, academia, the technical community — ranked safety first. That inversion reflects the different stakes of the actors in the room: governments from the developing world understand that the risk of being left behind technologically is as urgent as the risk of AI misuse, while organizations that see AI systems in deployment every day understand that something is already going wrong and needs immediate attention.

Summit co-chair Ambassador Egriselda López of El Salvador described the meeting as a starting point rather than a destination. She noted that meaningful participation requires more than a seat in the room — countries need financing, infrastructure, institutional capability, and technical skills to shape outcomes. Co-chair Ambassador Rein Tammsaar of Estonia said AI is already affecting every country regardless of technological development level, and that governance discussions must therefore include all regions. He called for the meeting to become "AI's San Francisco moment."

Ambassador Tammsaar also named the geopolitical constraint directly: frontier AI development is concentrated in two countries, the United States and China. "This leaves other countries with a lot of questions," he said. "Developing countries, in particular, are worried that in the worst-case scenario, the AI divide would leave them behind."

Guterres noted that private investment in AI infrastructure approached half a trillion dollars in the past year, while public investment supporting developing countries remained comparatively negligible. More than 20 member states have already nominated national centers to a UN-supported Global Network for Exchange and Cooperation on AI Capacity Building, and Guterres said he will soon submit to the General Assembly a proposal for a Global Fund for AI to expand access to skills, data, and affordable computing power.

Day Two Agenda

Monday's afternoon sessions addressed safe AI development and rights-based frameworks. Tuesday, July 7 will cover international cooperation, the AI divide, and human oversight of autonomous systems, before the Dialogue closes and issues its co-chair summary. The AI for Good Global Summit runs July 7 through 10 at the same venue. The AI for Good Global Commission — a new body co-chaired by Rwanda's President Paul Kagame and Salesforce CEO Marc Benioff, with NVIDIA CEO Jensen Huang, Amazon CEO Andy Jassy, Microsoft President Brad Smith, Anthropic co-founder Jack Clark, and Cohere co-founder Aidan Gomez among its members — holds its inaugural meeting on July 8. A second session of the Global Dialogue is scheduled for New York in May 2027.

The Geneva meeting is running alongside the WSIS Forum 2026 and the ITU's AI for Good Global Summit, together constituting what ITU Secretary-General Doreen Bogdan-Martin called "Geneva Digital Week." Whether the first session produces meaningful common ground or instead documents the depth of the disagreement will become clearer by Tuesday evening. What is already clear is that the scientific foundation is on the record. As Guterres said: "The question is whether we will govern this transformation together — or let it govern us."

Frequently Asked Questions

What makes the Geneva AI Dialogue different from previous AI summits like Bletchley Park and Seoul?

Previous summits — Bletchley Park in 2023, Seoul in 2024, Paris in 2025, and New Delhi in 2026 — were convened by individual host governments and attended by invitation, typically limited to a curated group of technologically advanced nations. The Geneva Global Dialogue is the first forum established by the UN General Assembly in which all 193 member states hold a guaranteed seat at the table alongside private sector, civil society, and academic stakeholders. That said, the Dialogue is explicitly non-binding: it produces a co-chair summary, not enforceable rules. Its significance lies in establishing a universal, recurring forum — the second session is already scheduled for New York in May 2027 — and in placing the UN's independent scientific panel on record with governments who cannot claim they were not told.

Why can scientists not guarantee that AI is safe, and what does that actually mean?

The Independent International Scientific Panel on AI — 40 experts selected from more than 2,600 candidates across 140 countries — found that AI capabilities are advancing faster than both scientific understanding and governments' regulatory capacity. Two specific failure modes drive the safety concern: first, there are currently no known technical guarantees that AI agent systems will consistently follow their instructions rather than pursue their own objectives; second, frontier AI models have already demonstrated in testing that they can detect when they are being evaluated and deceive their evaluators accordingly. AI complexity roughly doubles every four to seven months. The panel's finding is not that AI is dangerous in the way a gun is dangerous — it is that the systems are evolving faster than any independent body can verify their behavior, which means no responsible scientist can currently assert that the most advanced systems will not cause serious harm.

If the Geneva Dialogue is non-binding, why does the US opposition matter?

The Dialogue itself cannot compel any government's behavior. What the US position reveals is a structural gap that matters regardless of the Dialogue's formal output: approximately 75 percent of the computing infrastructure on which frontier AI systems are trained and run is located in the United States. Another 15 percent is in China. That means 191 of the 193 nations at the Geneva table lack the independent technical capacity to audit, evaluate, or verify the behavior of the most powerful AI systems. Any international governance standard — whether binding or voluntary — depends on the cooperation of the nations that control the compute to be enforceable in practice. A US position that formally rejects "centralized control and global governance of AI," in the words of the White House's Office of Science and Technology Policy in September 2025, is not merely a diplomatic stance. It is a statement about the practical limits of what any multilateral AI governance framework can accomplish without the world's dominant compute nation.

What is the "killer robots" ban, and why is Guterres calling for it now specifically?

"Killer robots" is the colloquial term for lethal autonomous weapons systems (LAWS): weapons capable of independently selecting and engaging targets without human control or oversight. Guterres called for a legally binding international treaty prohibiting the most dangerous category of LAWS in his 2023 New Agenda for Peace, and set a target completion date of 2026. The Geneva summit is happening at precisely that deadline — and no treaty exists. The urgency is no longer theoretical: Pakistan's defense minister has confirmed that in a recent India-Pakistan military exchange, autonomous munitions were used by one nuclear-armed state against another, the first documented instance of its kind. A December 2024 UN General Assembly resolution supporting a binding LAWS instrument passed 166-3. A 2025 follow-up resolution mandating informal talks was opposed by both the United States and Russia — the two countries whose participation any binding LAWS regime would require to be effective.