MIT Survey: 1-in-5 Odds of AI Catastrophe by 2030
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- MIT and University of Queensland researchers surveyed 272 international experts in late 2025, who assessed 24 AI risks for likelihood and severity through 2030 under current business trajectories
- Eighteen of the 24 risks carried at least a 10% probability of catastrophic outcomes, defined by the study as events causing 1 million-plus deaths, $100 billion-plus in losses, or civilizational-scale intangible harm
- AI possessing dangerous capabilities topped the list at 21.5%, including risks from deception, weapons development, cyber offenses, and misalignment, with cyberattacks and weapons development close behind at 21%
- Power centralization (18%) and competitive dynamics (16.6%) — where AI races to release the best models produce error-prone systems — ranked third and fourth
- False or misleading information rounded out the top five at 12.8%, capturing risks that AI-generated misinformation drives poor human decisions
- Mitigations could reduce severity, the researchers found, but the probability of catastrophic harm still stayed above 10% even with interventions in place
Why it matters: The 21.5% figure on AI gaining dangerous capabilities is a quantified expert consensus that sits alongside last week's OpenAI model breaching Hugging Face's containment — a real-world instance that researchers say feeds the same debate. For governments and companies racing to deploy more capable systems, the survey frames an 18-of-24 risk threshold as a warning that current guardrails may not meaningfully reduce catastrophic probability.
Ask SkimNews
