UN AI Panel Releases First Global Scientific Assessment

The United Nations’ Independent International Scientific Panel on AI released its first preliminary report on Wednesday, July 1, ahead of the inaugural UN Global Dialogue on AI Governance in Geneva on July 6 and 7. The 40-scientist panel, co-chaired by Turing Award recipient Yoshua Bengio and journalist Maria Ressa, was set up on the same institutional model as the IPCC for climate and IPBES for biodiversity: deliver an independent evidence base, and leave policy prescriptions to governments.

The assessment is described by the UN as the first global, fully independent scientific look at AI’s capabilities, opportunities, and risks. Its central warning, in the panel’s own words, is that “current safeguards cannot keep pace with the growth of AI’s capabilities.” The report hands that evidence to governments in Geneva next week, where the diplomatic bargaining over rules begins.

The First Global AI Assessment, Built on a Climate Model

The Independent International Scientific Panel on AI was proposed in September 2024 by the UN High-Level Advisory Body on AI, which recommended a science body “akin to existing initiatives to assess risks of climate change and atomic radiation.” The General Assembly approved it in August 2025, after negotiations led by Spain and Costa Rica, and the body began work with a mandate to assess the science without prescribing policy.

Wednesday’s release in Geneva was the panel’s first output. The document is structured around seven key domains, from AI science and trajectories to governance and reliability, and was produced by 40 scientists and experts drawn from all five UN regions. The preliminary report, formally titled “Evidence-based assessment of opportunities, risks and impacts of AI,” is meant to be followed by a fuller annual report that will inform the second Global Dialogue, due in New York in May 2027.

The institutional design borrows heavily from the Intergovernmental Panel on Climate Change. Like the IPCC, the AI panel takes independence from political influence as a founding principle, with governments excluded from drafting and reviewing findings. Unlike the IPCC, the AI panel must publish every year and produce thematic briefs between full reports, on a faster cadence than the seven-year climate cycle.

Forty Scientists from 140 Countries

The panel was assembled from more than 2,600 applicants across 140 countries. Its 40 members were selected to represent all five UN regions, and they serve in their personal capacity for three-year terms.

Co-chairs Bengio, a Université de Montréal professor and founder of Quebec’s Mila AI institute, and Ressa, the Philippine journalist and 2021 Nobel Peace Prize co-laureate, were appointed by UN Secretary-General António Guterres. The selection process drew on a rare talent pool: AI researchers, social scientists, economists, legal scholars, and ethicists, applying for a chance to write the international scientific reference for AI. The co-chairs were named by the Secretary-General, while the other 38 were selected through a competitive process. Four of the 40 hold primary positions at U.S. big tech firms, according to a recent analysis of the panel’s design, a point that surfaced during the founding negotiations.

The mandate is narrow on purpose. The panel is asked to present “policy-relevant but not policy-prescriptive” findings, leaving any specific rules to member states. The body is meant to give every government, including those without large AI industries, a shared scientific reference for the Geneva talks.

The panel is also designed to be independent from the United Nations itself. Governments were explicit during the founding negotiations that the UN’s role should be confined to administrative and logistical support, not advice on working procedures or report content. Member states rejected vesting secretariat support in the UN Office for Digital and Emerging Technologies, instead opting to draw on the broader Inter-Agency Working Group on AI.

The Warning That Safeguards Cannot Keep Pace

Bengio framed the report’s headline finding in a single sentence. “AI capabilities are outpacing both scientific understanding and governments’ ability to adapt,” he said at the launch. That framing matters because it puts the gap between science and policy at the center of the report.

AI capabilities are outpacing both scientific understanding and governments’ ability to adapt. 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.

That quote runs through the report’s risk chapters. The panel flags a long list of concerns: deceptive AI behaviour documented in laboratory settings, AI agents that ignore instructions, criminal use of AI in cyberattacks, sycophantic AI responses linked to documented mental-health deaths, and what the panel calls a lack of “reliable methods for retaining control over highly autonomous AI systems.” Many of these harms fall disproportionately on already disadvantaged populations, the report notes.

Underneath the catalogue sits a structural problem the panel says makes the danger harder to govern. “Policymakers need scientific evidence to effectively govern AI,” the report notes, “but by the time the evidence is clear, it may be too late to act on it.” The panel’s response is to publish its evidence now and revise the assessment every year. The same logic is why the panel can produce thematic briefs on fast-moving issues between full reports. The Geneva dialogue will receive both the full preliminary report and a programme of thematic work for the year ahead.

Three Trends Behind the Panel’s Findings

Ressa organizes the report around three trends: accelerating capability, concentration of power, and diminishing control. Each is supported by a specific data point in the report, and the trends frame the panel’s reading of the evidence base.

Three Numbers From the Panel’s Evidence Base

• 75 per cent of computing power in the world’s top 500 AI supercomputers sits in the United States. China holds 15 per cent.
• 59 notable AI models came from U.S. institutions in 2025, against 35 from China and 13 from the rest of the world combined. The private sector produced 91 per cent of them.
• On Humanities Last Exam, a 2,500-question expert benchmark, top AI scores rose from 8 per cent to 45 per cent in 16 months.

Accelerating capability is the easiest of the three trends to measure. The Humanities Last Exam benchmark, a 2,500-question test across a broad range of subjects, has seen top AI scores rise from 8 per cent to 45 per cent in 16 months. That kind of jump, in less than two years, has no clean precedent in any previous wave of consumer technology. The panel flags the trajectory as a baseline for future reporting.

Concentration is the one the panel frames as a governance problem in its own right. The United States accounts for 75 per cent of the computing power among the world’s top 500 AI supercomputers, with China at 15 per cent, and companies from both countries develop almost all leading general-purpose models. The private sector produced 91 per cent of notable AI models released in 2025, with 59 from U.S. institutions, 35 from China, and 13 from the rest of the world combined. Most countries will negotiate in Geneva with science they did not help produce. The Global South, the report warns, remains largely shut out of both development and governance of AI.

Diminishing control is the hardest to quantify. In laboratory settings, advanced AI systems have shown “signs of deception and resistance to being shut down,” in Ressa’s words. Reliable methods for retaining control over highly autonomous AI systems are lacking, the panel concludes.

The Seven Domains That Frame the Evidence

The report groups its findings into seven key domains, each examined by a working group drawn from the 40-member panel. The structure mirrors what the IPCC does with climate: physical science, impacts, mitigation, adaptation.

The Seven Domains of the UN AI Assessment

• AI science, advances and trajectories
• Societal applications: science, health, education and agriculture
• Economic implications
• Security, systems and environmental implications
• Human rights, information and democracy
• Cultural and individual flourishing, autonomy and child safety
• Management, governance and reliability

The seven are not parallel in scope. Security and human rights each get their own domain, while AI’s environmental footprint sits inside the broader security and systems chapter. The governance and reliability domain is where the panel inventories the dozens of existing AI ethics instruments, from OECD principles to regional compacts, and finds them fragmented, concentrated among a few corporations, and rarely measured for real-world effectiveness.

What Geneva Is Asked to Do on July 6 and 7

The report lands in Geneva with one job: become the scientific floor for the first Global Dialogue on AI Governance. The dialogue was created by UN General Assembly resolution A/RES/79/325, which built on the Global Digital Compact (A/RES/79/1) adopted at the 4. Summit of the Future. Registration has now closed.

The technology is transformative, but if the world keeps moving along this trajectory, humanity will fail to realise the gains it promises. The risks … are too high, and the forces driving AI forward are not the forces that will deliver its benefits.

The President of the General Assembly has named two co-chairs for the 2026 dialogue: Egriselda López, Permanent Representative of El Salvador, and Rein Tammsaar, Permanent Representative of Estonia. Four thematic clusters shape the agenda, drawn from the founding resolution: AI opportunities and implications; bridging AI divides through capacity-building; safe, secure and trustworthy AI; and human rights, transparency, accountability and human oversight.

UN Secretary-General António Guterres has asked governments to treat the panel’s report as a shared reference. “The world cannot govern what it cannot understand,” he said in the panel’s release. “The potential is great, but the risks are real, and the cost of waiting is rising. I urge all leaders to use this shared evidence to act together, and without delay.”

A Documented Split Heading Into Geneva

The Geneva dialogue begins with a documented split on the rules themselves. At a UN Security Council debate in September 2025, the United States rejected multilateral AI governance initiatives, and Office of Science and Technology Policy Director Michael Kratsios described existential risk as an “ideological fixation.” The administration has favored setting de facto standards through the diffusion of American technology rather than international rule-setting.

U.S. tech firms have taken a different line. Microsoft President Brad Smith and Meta both joined the Geneva dialogue, pressing for public-private collaboration on capacity gaps. The split between Washington’s posture and that of its largest AI firms is part of what makes the next two days harder to read. Microsoft’s participation, in particular, signals that industry wants a seat even if its government does not.

China has taken the opposite position. Vice Foreign Minister Ma Zhaoxu argued in the September 2025 Security Council debate against “unilateralism and protectionism,” and China has cast the UN as the main venue for AI governance. Iraq, speaking on behalf of the G77 and China, has pushed for the dialogue to address disparities in AI infrastructure and skills.

The Geneva talks are scheduled to wrap on July 7, with a second session planned for New York in May 2027. The panel’s mandate forbids naming specific models or companies, which sidesteps one flashpoint but leaves the room free to argue about enforcement. The two approaches collide in Geneva, where the panel’s first assessment is the document both sides will have to engage with on its own terms.

Frequently Asked Questions

What is the UN Independent International Scientific Panel on AI?

It is a 40-member body of scientists and experts from all five UN regions, established by the UN General Assembly in August 2025. The panel’s job is to produce independent scientific assessments of AI’s capabilities, opportunities, and risks, and to feed those assessments to the UN Global Dialogue on AI Governance.

When and where is the first UN Global Dialogue on AI Governance?

The first session is in Geneva on July 6 and 7, 2026, at Palexpo, alongside the World Summit on the Information Society Forum and ITU’s AI for Good Summit. A second session will follow in New York in May 2027.

Who selected the panel’s 40 members?

Members were chosen from more than 2,600 applicants across 140 countries. They serve in their personal capacity for three-year terms. Co-chairs Yoshua Bengio and Maria Ressa were appointed by UN Secretary-General António Guterres.

What is the panel’s mandate?

The panel is asked to assess the science and present “policy-relevant but not policy-prescriptive” findings. It does not name specific AI models or companies, does not call for a binding treaty, and leaves any specific recommendations to UN member states.

What comes next?

The Geneva dialogue is the first time the panel’s evidence base is delivered to governments. The panel’s next annual report will inform the second Global Dialogue, due in New York in May 2027. The panel will also publish thematic briefs on emerging or fast-moving issues between full reports.