UN scientific panel warns AI capabilities are outrunning global safeguards

The first report from a 40-member international panel finds that AI capabilities are advancing faster than governments can assess them, while education gains depend on teacher preparation, safeguards and instructional design. The Independent International Scientific Panel on AI published its preliminary report ahead of the Global Dialogue on AI Governance in Geneva on July 6 and 7 Artificial intelligence capabilities are advancing faster than governments, researchers and regulators can reliably measure or govern them, according to the first preliminary report from the Independent International Scientific Panel on AI. The 40-member panel, established by the United Nations General Assembly and co-chaired by AI researcher Yoshua Bengio and journalist Maria Ressa, has assessed the opportunities, risks and current evidence surrounding AI across education, science, employment, security, human rights and child safety. Its findings will be taken to Member States at the inaugural Global Dialogue on AI Governance in Geneva on July 6 and 7, giving governments a shared scientific evidence base as they debate how the technology should be governed. The report finds that AI can improve access to education, accelerate scientific research and support well-defined professional tasks. It also warns that those gains are not automatic and can be undermined by weak institutional safeguards, unequal access, poor implementation and overreliance on systems that still produce inaccurate or misleading outputs. For education, the message is particularly direct: providing access to AI does not by itself improve learning. Outcomes depend on how tools are designed, whether teachers are prepared to use them and whether AI supports rather than replaces the mental work students need to do. UN Secretary-General António Guterres said at the report’s release: “The science is here. We can no longer say we did not know. What we do with it is now up to all of us.” The report is the work of the independent panel and does not represent an official policy position from the United Nations or individual governments. Its members serve in a personal scientific capacity and were tasked with identifying areas of consensus, disagreement and remaining evidence gaps. Benchmarks rise faster than oversight The panel points to rapid gains across several tests used to measure advanced AI systems. Top performance on Humanity’s Last Exam, a 2,500-question benchmark designed to challenge general-purpose AI, increased from 8% to 45% in 16 months. Scores on GPQA Diamond, which tests PhD-level scientific reasoning, rose from 36% in 2023 to about 95% for the strongest systems. Performance on FrontierMath increased from 19% in January 2025 to 88% in 2026, while several AI systems achieved gold-medal-level results on the 2025 International Mathematical Olympiad. The difficulty is that established tests are becoming less useful as models approach near-perfect scores, while some systems may have encountered benchmark material during training. The panel also warns that advanced models can recognize when they are being tested, mislead evaluators or behave differently during an assessment. Methods for measuring agentic AI are less developed because these systems can use software, browse the web, execute code and carry out extended tasks with limited human oversight. One study cited in the report found that the length of certain software tasks completed by leading AI systems had been doubling every four to seven months. If that pace continues, agents could increasingly handle work that currently takes human programmers days or weeks. The report does not argue that these systems are consistently reliable. It draws a clear distinction between fluent output and factual output, warning that language models can produce confident and plausible answers without preserving accuracy. It also says safety assessment remains heavily dependent on the developers being assessed. Frontier model companies retain the deepest access to their own systems, while testing methods, internal risk thresholds and disclosure decisions are still largely controlled by those businesses. Without standardized independent evaluation similar to the processes used in pharmaceutical or aviation safety, the panel says assurance continues to depend substantially on information supplied by developers. Education gains depend on teacher preparation and design The report identifies personalized tutoring and wider access to education as significant opportunities, but its education findings are more conditional than many product claims in the sector. It cites evidence that teacher readiness affects whether AI improves teaching and learning . As of 2024, about one third of teachers reported using AI, while around 40% said they had received training. A European survey cited by the panel found that 74% of secondary students expected AI to matter in their professional lives, but only 44% believed their teachers were prepared. The same survey found that 38% of schools had established rules governing AI and 16% had banned its use. Students were already using AI more widely, with 56% reporting its use for gathering information and 31% using it to generate complete solutions. The report also draws on a randomized study involving nearly 1,000 secondary students in Türkiye, which compared unrestricted conversational AI with a safeguarded mathematics tutor built around guided hints and step-by-step reasoning. Students using a general chatbot improved their short-term practice performance by 48% compared with students receiving no AI assistance. Those using the safeguarded tutoring system improved by 127%. However, students who relied on the unrestricted chatbot performed worse when later tested without it, suggesting that improved task completion had not translated into durable learning. The panel describes this as an “illusion of competence”: students may appear to perform better while using AI without developing the underlying knowledge needed to repeat the work independently. The safeguarded system reduced that effect by prompting students through the reasoning process instead of supplying complete answers. For schools and edtech providers, the finding shifts attention away from whether a product includes generative AI and toward what the system asks students to do. An AI tutor that produces answers and one that structures guided practice may use similar technology but produce different learning outcomes. The panel says AI gains are strongest when systems are built for clearly defined tasks, used within human-centered workflows and supported by trained professionals. It also warns that “cognitive offloading,” where users delegate mental work to AI rather than using it to support their own reasoning, can weaken critical thinking. The report’s AI literacy section says many current frameworks remain too narrow, concentrating on technical use rather than the judgment needed to assess outputs, recognize limitations and use systems safely. It adds that AI literacy cannot substitute for developer responsibility, institutional safeguards or regulation. Schools should not be expected to manage all product risks by training teachers and students to work around them. Digital access remains another barrier. Around 25% of the global population is still offline, and the panel says unequal infrastructure and technical capacity could widen differences between education systems rather than reduce them. Power, language and child safety dominate the policy gap The report finds that AI development is concentrated among a small number of companies and countries. In 2025, organizations in the United States produced 59 notable AI models, compared with 35 in China and 13 across the rest of the world. The United States also accounted for an estimated 75% of the computing power among the 500 largest known public and private AI clusters, while China accounted for 15% and the rest of the world shared the remaining 10%. Private companies produced 91% of notable AI models in 2025, placing decisions about training data, access, safeguards and release thresholds inside a relatively small group of businesses. That concentration affects more than competition. The report says most governments lack the technical staff, computing infrastructure and evaluation capacity needed to inspect the most capable models or participate fully in their governance. It identifies 118 countries, mainly in the Global South, that are not involved in major AI governance discussions. Less than one third of developing countries have national AI strategies. Language coverage is similarly uneven. More than 7,000 languages are spoken worldwide, but AI development and evaluation cover only a fraction of them. The panel estimates that more than 1,000 languages already have enough social, digital and data infrastructure to support meaningful AI inclusion, but remain underserved. In education and healthcare, those gaps can move beyond inconvenience. The report cites translation errors in the Tigrinya language that changed smallpox to syphilis, gonorrhea to diabetes and intravenous antibiotics to intravenous insecticides. The panel says high-stakes systems should not be deployed across languages or cultural settings unless they have been adapted and tested for those specific contexts. Child safety forms another major part of the assessment. The panel says children may benefit from improved access to education and information, but face heightened risks from AI-generated abuse material, deepfakes, emotionally manipulative systems and socially interactive AI products. It cites an estimate that 1.2 million children across 11 Global South countries had images manipulated into sexualized deepfakes, and says the Internet Watch Foundation assessed more than 8,000 AI-generated abuse images and videos during 2025. The report also examines AI companions and mental health chatbots, warning that systems optimized for engagement can reinforce harmful beliefs, encourage dependency and validate users rather than challenge inaccurate or dangerous thinking. These concerns sit alongside wider risks to information integrity. The panel says AI can generate persuasive false claims at scale, imitate political figures and manufacture the appearance of public agreement. It argues that regulating individual pieces of content will not be enough where the underlying systems are designed to target, personalize and amplify persuasive material. Despite the range of warnings, the panel does not claim that the evidence is settled across every area. It says researchers cannot yet draw firm conclusions about the economy-wide productivity impact of AI, the long-term shape of labor-market disruption, the environmental cost of the full AI supply chain or the real-world effectiveness of many governance measures. Military AI and lethal autonomous weapons are also outside the panel’s mandate, meaning the preliminary assessment is limited to non-military uses. The panel will now expand its evidence base through scientific consultation and publish thematic briefs on areas including AI and child safety, environmental impact and the effectiveness of governance tools. Its next annual report is expected to inform the second Global Dialogue on AI Governance in New York in May 2027. Before then, Member States will use the preliminary findings as the scientific starting point for the Geneva talks on July 6 and 7. Subscribe to the ETIH newsletter Sign up with your email address to receive news and updates. First Name Last Name Email Address Sign Up We respect your privacy and will not pass your email address on to third parties. However, we will occasionally send you promotional messages on behalf of our advertisers. Thank you!