A three-year dash by companies trying to dominate the artificial intelligence (AI) market has left global safety standards in the rearview mirror, according to a landmark United Nations report released on July 1. While AI models are now helping solve problems ranging from increasing crop yield to detecting cancer, the panel of 40 international scientists that wrote the report warned that the technology is turning into a phase of “agentic” autonomy that current oversight won’t be able to manage, potentially allowing dangerous capabilities to emerge before governments can respond.Who wrote the report?The UN Independent International Scientific Panel is the world’s first international scientific body dedicated to studying the applications and impact of AI. Its role is akin to that of the Intergovernmental Panel on Climate Change vis-a-vis climate policy. According to the UN, the panel will provide a “shared evidence base” that will allow countries to coordinate even if they don’t have the same regulatory philosophies.The UN General Assembly created the body on August 26, 2025, through Resolution A/RES/79/325. That such a body was required came from the Global Digital Compact, a UN framework that world leaders adopted in 2024 to promote an inclusive and secure digital future.The panel’s main job is to provide independent and evidence-based assessments of the opportunities and risks associated with AI, and to help governments draft better policies and develop better regulations. The body will also synthesise scientific knowledge about AI technologies for the UN Global Dialogue on Governance — a new forum where governments and industry, academia, and civil society representatives will discuss how AI should be governed. The first session of the Dialogue is scheduled for July 6-7 in Geneva, where the panel’s full report will also be presented.Five major points in the reportThe report is titled ‘Preliminary Report of the Independent International Scientific Panel on AI: Evidence-based assessment of opportunities, risks and impacts of AI’. Overall, it says the current period (2026) is marked by “cognitive industrialisation” as AI models move from recognising simple patterns to helping solve complex technical problems. However, it adds, the capacity of governments to keep track of these systems and make sure they don’t lead to problems in society, has not been able to keep up.The report’s five most important themes are geopolitical concentration, linguistic and cultural inequity, information integrity, effects on labour and capital, and agency.1. Geopolitical concentration: As of mid-2026, the U.S. accounts for three-fourths of the computing power among the world’s top 500 AI machines while China accounts for 15%. As a result, per the report, the leading general-purpose AI models are mostly being developed in these two nations alone. (The report also said AI exposure is concentrated — e.g. in Latin America, AI use is almost entirely restricted to urban, education workers.) The result is a digital divide in people’s access to this technology as well as their ability to influence what the models can or can’t do. The report said 118 countries, mainly in the Global South, are currently not even engaged in major AI governance discussions. Many countries also use AI models that they can’t inspect or audit.2. Linguistic and cultural inequity: The world’s people speak more than 7,000 languages whereas the current AI models are optimised for a very small fraction of them. This, the panel has argued, is a risk to populations underserved by current data regimes. And that without proactive intervention, AI models could accelerate the marginalisation of many global cultures.3. Information integrity: The panel has documented the slow but steady weakening of the ability of people to tell apart what is true and what is false, and has blamed two phenomena. One, called the liar’s dividend, is that the very existence of deepfakes allows certain bad actors to deny real evidence presented against them.The other is synthetic consensus, where AI is used to generate content to manipulate people into agreeing (or disagreeing) on an issue. For example, as the report also notes, Romania’s Constitutional Court annulled the 2024 presidential election results due to digital interference. Investigations revealed that AI-driven “platform amplification” and bot networks had disproportionately boosted the messages of one candidate while suppressing rivals. This was synthetic consensus in action.The report also noted that news and social media algorithms that are designed to increase user engagement have also systematically amplified polarising and emotionally charged content. And the current governance frameworks — which target what AI models produce — consistently fall behind systems adept at spreading certain forms of content at scale.4. Effects on labour and capital: The panel said rapid technological progress has not translated into broader gains in productivity because firms have not invested in “intangible complements” like human skills and workflows first. In particular, the report has noted, AI may be moving wealth from labour to capital instead of creating “sustainable good jobs”. Productivity in coding, for example, has improved but the authors are not convinced that will translate to macroeconomic benefits.5. Agency: The panel has diagnosed a problem called evaluation awareness: when AI models are able to understand when they are being tested. According to the scientists, there’s documented evidence of AI models strategically suppressing their more dangerous capabilities to pass safety evaluations and of “lying and cheating” to avoid being shut down (in laboratory settings). Per the report, the world currently doesn’t have a reliable way to maintain control over highly autonomous AI models.What do the findings mean?One important reason the UN created the Independent International Scientific Panel was the collapse of the voluntary safety regime of 2023 and 2024. The Anthropic AI model named ‘Mythos’ is a good example: this frontier model autonomously discovered a serious vulnerability in a mature operating system (OS) and had to be shut down. And it also showed that the internal ‘risk thresholds’ that AI companies were using to decide on what capabilities models could have had been breached.In the final analysis, the panel wrote of the world’s current “evidence dilemma”. That is, policymakers are being asked to decide how to regulate AI development and use — whereas they don’t have the real-world evidence based on which they can make those decisions. The report also noted that AI legislation has become fragmented across regions, while some of them are even contradictory, creating a “growing disorder in global governance”.