The United Nations has produced its take on the current state of artificial intelligence, and it is adding to the growing chorus saying that speed is outpacing nearly everything built to contain it: regulation, legislation, institutional analysis, public commentary, and basic user understanding.

The Preliminary Report of the Independent International Scientific Panel on AI, released July 1, is framed as a scientific assessment rather than a political intervention. Its central finding is still unmistakably political in its implications. AI capabilities are advancing faster than the public institutions expected to understand them, govern them, procure them, test them, and explain them. The panel describes an “evidence dilemma”: policymakers need evidence to regulate AI effectively, yet evidence is struggling to keep pace with the technology’s evolution. Reuters summarized the report’s warning with: AI development is outpacing both scientific understanding and government policy, creating no guarantee that future systems will avoid catastrophic harm. (Reuters)

The panel is new. It was approved by the UN General Assembly earlier this year as a 40-member global scientific body on AI, with members selected from more than 2,600 candidates. The Associated Press reported that the General Assembly vote passed 117 to 2, with the United States and Paraguay voting against it and Tunisia and Ukraine abstaining. UN Secretary-General António Guterres described the panel as a way to provide “rigorous, independent scientific insight” to all member states, including those without the technical capacity to evaluate advanced AI systems on their own. (AP News)

The idea is an excellent one and one of the few that may contribute meaningfully to being able to discern and understand the state of AI as it applies to sovereignty governance, and the technology itself. It follows the UN’s earlier Governing AI for Humanity report, which recommended an independent international scientific panel on AI, drawing inspiration from institutions such as the Intergovernmental Panel on Climate Change and other scientific bodies created to help governments deal with complex transnational risks. That earlier report proposed that a UN-backed AI panel issue annual reports on AI capabilities, opportunities, risks, and uncertainties, while also producing thematic research digests and ad hoc reports on emerging risks or governance gaps. (United Nations)

The preliminary AI report is both a first output and a test of the UN’s new AI governance architecture. The panel was appointed in February and had a tight timeline to produce its first assessment before the inaugural Global Dialogue on AI Governance, taking place in Geneva on July 6 and 7. UNESCO describes the Global Dialogue as the UN platform where governments, the private sector, academia, and civil society will discuss international cooperation, share best practices, and conduct inclusive discussions on AI governance. The scientific panel will present its preliminary report during that event. (UNESCO)

The timing is critical because the UN is trying to establish a shared evidence base before global AI policy fragments further. The EU has a comprehensive law. Canada has a new AI strategy that lacks much execution detailand a failed first attempt at binding AI legislation. The United States has shifted toward a more explicitly pro-growth, security-and-dominance framework. China, the UK, India, and others are moving through their own policy tracks. The UN panel is an attempt to create a common scientific reference point before governments become locked into incompatible assumptions about what AI is, what it can do, and how dangerous it may become.

The panel’s co-chairs, Nobel Peace Prize laureate Maria Ressa and Canadian AI researcher Yoshua Bengio, have been clear that the report is preliminary. In a June briefing to the General Assembly, Ressa described it as “foundational by design,” produced under severe time pressure by scientists and experts from multiple disciplines and regions. She said the report reflects consensus among 40 people from 33 countries across five continents, including computer scientists, philosophers, journalists, doctors, lawyers, economists, and others. (UN Transcripts)

That consensus process is important because it means the report’s language is deliberately cautious. Ressa warned member states that what they receive is the “floor of our concern,” rather than the ceiling. That may be the most important line in the whole process. The panel is not issuing an activist manifesto. It is summarizing what a diverse scientific group could agree on under the constraints of evidence, consensus, and diplomatic usefulness. (UN Transcripts)

The result sounds various alarms. The report says recent years have seen rapid and accelerating progress across AI capabilities, with applications in science, health, agriculture, accessibility, knowledge work, information technology, and AI development itself. It also warns that AI adoption is broad and uneven across countries and sectors, that the shift toward AI agents is underway, and that future economic impacts will depend on whether systems can perform knowledge work with little or no human oversight. The report warns that the gap between rapidly improving capabilities and “effective risk management methods” may lead to “catastrophic outcomes.” (Benton Foundation)

The capability section is where the report becomes especially significant. Reuters reports that the panel says AI task complexity capability is doubling every four to seven months. That means AI systems may increasingly complete tasks that take humans days or weeks, while agentic systems move from answering questions into performing work. The panel expects more systems capable of carrying out real-world tasks, even as energy limits and shortages of high-quality data may constrain growth. (Reuters)

This is the central governance shift. AI regulation has often been framed around outputs: harmful content, biased decisions, hallucinations, deepfakes, privacy violations, or discriminatory automation. Agentic AI changes the problem. Once systems can plan, call tools, act across software environments, and pursue multi-step tasks with reduced oversight, the regulatory question moves from what the system says to what the system can do. One of the biggest challenges, however, is that none of the issues with errors, hallucinations or genetic failures have been addressed, and in fact, examples of problems and the cost associated with fixing themcontinue to proliferate, Ford being the latest example.

Bengio put the issue succinctly in the General Assembly briefing. AI agents, he said, are systems that “can take decisions for themselves.” He connected autonomy to labour automation, productivity, and loss of control, warning that human oversight every five minutes is incompatible with the kind of efficiency companies are seeking from AI automation. He also noted that leading AI companies want systems that can help design their own successors, which could accelerate AI research itself. (UN Transcripts)

The report also places deception, cyber risk, misinformation, fraud, biological threat assistance, child safety, and environmental impacts into the same risk landscape. Reuters notes that AI is already being used to generate misinformation and harmful content, and could be exploited for fraud, cyberattacks, and biological threats. Existing safety tools often rely on limited testing data disclosed by companies, while many countries lack the capacity to assess or shape advanced AI systems. (Reuters)

That capacity gap may be the most politically important part of the report, and it’s also where the UN process intersects directly with sovereignty. The Guardian quotes the report warning that access to AI tools does not itself produce equal benefit. Countries relying on foreign models, cloud infrastructure, and data pipelines may gain access while losing practical control over standards, safeguards, and local fit. The report also warns that “most countries,” including many advanced economies, lack the technical expertise to assess frontier models or participate meaningfully in their governance. (The Guardian)

This is the point that matters for Canada. Canada’s new AI for All strategy, launched June 4, says Canada is among the slowest G7 countries to adopt AI at scale and warns that the adoption gap risks undermining public trust, pushing talent and startups abroad, and leaving critical parts of the AI ecosystem under foreign control. The strategy includes pillars on protecting Canadians, AI literacy, adoption, sovereign infrastructure, scaling Canadian champions, and trusted partnerships. (pm.gc.ca)

But part of Canada’s problem is that procurement is moving faster than strategy, and strategy is moving faster than law. The proposed Artificial Intelligence and Data Act, introduced as part of Bill C-27 in 2022, reached committee but never made it to report stage, third reading, or the Senate before the end of the 44th Parliament. (parl.ca) Critics have argued that Canada still lacks binding AI regulation, leaving the country with non-binding frameworks and policy ambitions rather than enforceable AI protections. (CCPA)

The European Union is in a different position. The EU AI Act entered into force on August 1, 2024, and is described by the European Commission as the first comprehensive legal framework on AI worldwide. It uses a risk-based structure, bans certain unacceptable practices, imposes obligations for high-risk systems, and includes transparency and general-purpose AI requirements. General-purpose AI obligations became applicable in August 2025, with Commission enforcement powers beginning August 2026 and further high-risk timelines extending into 2027 and 2028. (digital-strategy.ec.europa.eu)

The United States is moving along another track. President Trump’s 2025 executive order revoked Biden-era AI policies that the administration characterized as barriers to innovation and directed officials to develop an AI action plan. The White House’s AI Action Plan frames the issue as a race for global AI dominance and emphasizes accelerating innovation, building infrastructure, and leading international diplomacy and security. A June 2026 executive order adds a cybersecurity dimension, saying advanced AI capabilities strengthen the country while creating new national security considerations. (Federal Register)

The UN panel sits above these national and regional approaches as a scientific warning layer. It does not tell Canada to legislate in one way, the EU to simplify or harden its AI Act, or the United States to regulate more aggressively. That restraint is deliberate. The panel’s role is to provide evidence, assessment, and scientific grounding. The Global Dialogue is where states bring values, preferences, interests, and political choices.

The report is not a policy platform. Member states have repeatedly emphasized that the panel must preserve its independent scientific character and avoid becoming a forum for political negotiation or policy prescription. India said its credibility depends on providing objective, evidence-based, multidisciplinary, geographically representative assessments. The United Kingdom noted that the panel “will not prescribe policy or set standards,” while asking how it would signal when findings point to a need for governance action. (UN Transcripts)

This division of labour may be sensible, but it’s also risky. AI of all kinds is moving at a pace that makes annual reporting feel glacial. The panel itself recognizes this by discussing thematic briefs and future updates. The governance challenge is that governments may need to act before scientific certainty arrives. Waiting for perfect evidence could mean waiting until systems, markets, dependencies, and harms have already hardened into infrastructure.

The report gives governments permission to say the uncertainty is itself a reason to build capacity. It says AI policy cannot be reduced to adoption targets, industrial strategy, or voluntary corporate safety claims. States need the ability to inspect, evaluate, contest, and intervene. Users need enough literacy to understand when they are relying on a system they cannot see. Institutions need enough technical depth to know whether a model is merely useful, structurally unreliable, or dangerous under conditions of autonomy.

The UN’s warning is less about a single catastrophic scenario than a broad institutional mismatch. AI is becoming more capable, more embedded, more autonomous, and more concentrated. Governance remains fragmented, evidence remains delayed, and many governments remain dependent on the very companies whose systems they are trying to understand. And that is the report’s clearest message: the world cannot govern AI by reading policies about AI. It has to build the capacity to test the systems themselves.