272 experts ranked the most urgent AI risks, and dangerous capabilities top the list
In the MIT-led Delphi study, experts scored 24 risk domains: under business as usual, 18 carried a catastrophic-outcome probability above 10 percent.
In a study with 188 co-authors, run by the MIT AI Risk Initiative (MIT FutureTech) and the University of Queensland, 272 international AI experts scored 24 risk domains using the Delphi method, a structured multi-round expert elicitation. Each risk was assessed for likelihood and severity of harm between 2025 and 2030 under two scenarios: business as usual, and pragmatic mitigation. MIT Sloan published the results.
The five most urgent risks came out as: AI systems acquiring dangerous capabilities, competitive pressures eroding safety, AI-enabled weapons and cyberattacks, centralization of power and unfair distribution of gains, and the spread of false or misleading information.
The starkest finding is the gap between scenarios: under business as usual, 18 of the 24 domains carry a probability of catastrophic outcomes above 10 percent. The bar for catastrophe is not vague either: more than 1 million deaths, over 100 billion dollars in losses, or civilization-scale damage. With pragmatic mitigation the count drops to five: dangerous capabilities (12 percent), weapons and cyberattacks (12 percent), environmental harm (12 percent), inequality and unemployment (11 percent), and power centralization (11 percent). Mitigation does not zero the risk, but it narrows the field dramatically.
The experts flag information, national security and finance as the most exposed sectors. The core team behind the work is Alexander Saeri, Jess Graham, Michael Noetel, Peter Slattery and Neil Thompson. Slattery sums up the goal in one line: figuring out who needs to do what differently.
