The International AI Safety Report 2026: Scenarios, not scaremongering

The second [International AI Safety Report](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026), published in February 2026, is not a sequel that rehashes its predecessor. It is a deliberate narrowing of scope and an escalation of method. Led by Turing Award winner Yoshua Bengio and authored by over 100 AI experts, the report is backed by more than 30 countries and international organisations. That makes it the largest global collaboration on AI safety to date. But size is not the story.

The report introduces two structural changes that matter more than its headcount.

First, it narrows its focus to “emerging risks”: risks that arise at the frontier of AI capabilities. That is a defensible editorial choice. General-purpose AI produces a sprawling risk surface — bias, disinformation, labour displacement, environmental cost, concentration of power. A single report cannot cover all of it with rigour. By carving out frontier risks as its domain, the report signals where the uncertainty is highest and where the evidence base is thinnest. It also carves out space for complementary efforts, including the United Nations’ Independent International Scientific Panel on AI, to cover the rest.

Second, the report draws on new research from the OECD and the Forecasting Research Institute to present specific scenarios and forecasts. This is a meaningful departure from the first edition, which catalogued risks but did not formalise their likelihood. Policymakers need to know not just that a risk exists, but how probable it is under different timelines. The report now gives them that, within the limits of the science.

What the report does not do is endorse any particular policy or regulatory approach. The Expert Advisory Panel, comprising representatives from over 30 countries including China, the EU, India, Kenya, and the United States, provided technical feedback only. The report is a synthesis of existing research, not a political document. That is both its strength and its limitation.

The strength is credibility. When a report backed by Beijing and Washington, by the OECD and the UN, says something about AI risk, it cannot be dismissed as the product of one government’s agenda. The limitation is that synthesis without prescription leaves the hardest work to others. Knowing that frontier models pose a risk of catastrophic misuse is useful. Knowing which regulatory lever to pull, and when, is the question that keeps civil servants awake at night. The report does not answer it.

The report’s list of contributors reads like a who’s who of AI governance. Senior advisers include Daron Acemoglu, Geoffrey Hinton, Stuart Russell, Marietje Schaake, and Andrew Yao. The writing group draws from Stanford, MIT, Oxford, Cambridge, Princeton, Carnegie Mellon, and the Alan Turing Institute. Industry reviewers include Anthropic, Google DeepMind, Meta, Microsoft, OpenAI, and Hugging Face. The breadth of institutional buy-in is unprecedented.

But breadth can mask disagreement. The report does not resolve the deep splits in the AI safety community: between those who see near-term misuse risks as paramount and those who worry about long-term loss of control; between those who favour transparency and those who argue that openness enables harm; between those who think regulation should focus on compute and those who want to regulate model outputs. The report catalogues these tensions without settling them. That is honest, but it leaves policymakers without a clear path.

The most telling line in Bengio’s foreword is this: “The pace of AI progress raises daunting challenges. However, working with the many experts that produced this Report has left me hopeful.” Hope is a diplomatic word. It is also a hedge. The report’s scenarios almost certainly include timelines on which frontier models develop capabilities that outstrip existing control measures. It is hard to read that and feel hopeful unless hope means “we are finally taking this seriously enough to measure it.”

For AI builders, the report is a signal that the window for self-governance is closing. If the second edition is more rigorous and more specific than the first, the third will be more so. Governments that back this process are building a shared evidence base that will eventually underpin regulation. The question is not whether regulation comes, but whether it is informed by the science or driven by panic after a visible failure.

The report’s focus on emerging risks is also a reminder that the frontier is not static. A risk that is emerging today may be routine tomorrow. The report is an annual exercise. That cadence is both a commitment and a constraint. It keeps the evidence base current, but it also means that the report is always catching up to the labs. The next edition will have to contend with whatever capability jumps occur between now and early 2027.

The report is available in all six official UN languages. That is a practical choice with political weight. AI safety is not a rich-country concern. The risk of catastrophic misuse, of labour displacement, of democratic erosion — these do not stop at borders. A report that speaks to Nairobi, Brasília, and Riyadh in their own languages is a report that takes that seriously.

What the report does not do is name the specific models, the specific training runs, the specific deployment decisions that create the risks it describes. That is by design. The report is a synthesis of science, not a whistleblower complaint. But the abstraction can feel evasive. When a report warns of “emerging risks at the frontier” without naming the frontier, it leaves room for denial.

The report’s value is in the scaffolding it builds for future decisions. It establishes a shared vocabulary, a shared set of scenarios, and a shared set of forecasts. That is the prerequisite for coordinated action. The action itself is still to come.

What to watch next is whether the governments that endorsed the report act on its findings. Endorsement is cheap. Funding for AI safety research, staffing for regulatory bodies, and willingness to impose binding rules on frontier labs — those are the real measures. The report gives them the evidence. The rest is politics.