What the experts assessed

The researchers convened 272 specialists from industry, academia, government, and civil society across 37 countries. In three anonymous rounds, the participants assessed 24 AI risk categories under a business-as-usual trajectory and a scenario with pragmatic, cost-effective mitigation. The five-year horizon was intended to keep the exercise close enough for grounded judgment while capturing risks that are still emerging.

The study defined catastrophic harm as more than one million deaths, more than 100 billion dollars in financial loss, or comparable damage. Under business as usual, 18 categories received estimates above a 10 percent probability threshold. With pragmatic mitigations, five categories remained above it.

The number needs its caveat

These are elicited expert judgments, not observed frequencies. The categories overlap and may interact, so the researchers say they should not be added into a single overall probability. The panel also consisted of people specializing in AI risk, which makes their prioritization valuable while distinguishing it from a representative survey of every AI researcher or the public.

The strongest reading is therefore not that catastrophe is scheduled or that the probabilities are actuarially precise. It is that many informed participants judged a wide range of severe harms high enough to justify immediate mitigation, even when practical interventions were assumed.

Responsibility and exposure are separated

Dangerous model capabilities, AI-enabled weapons and cyberattacks, competitive dynamics, concentrated power, and sophisticated false information ranked among the most severe concerns. Information technology, finance, national security, healthcare, and education were identified as particularly exposed sectors.

The study also identifies a political imbalance. Developers, governments, regulators, and standards bodies were considered best positioned to reduce the risk, while users and the general public were expected to absorb much of the harm. That gap makes collective rules, independently measured mitigation, and public accountability more than optional safety practices.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

MIT Sloan — International AI experts warn of potentially catastrophic risks MIT FutureTech and University of Queensland — Prioritization of Risks from Artificial Intelligence