International AI experts warn of potentially catastrophic risks from AI

MIT FutureTech just released a new study “Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts,” in partnership with the University of Queensland School of Psychology. As part of this study, 272 international experts in AI risks from across the AI industry, academia, governments, and civil society, examined 24 categories of AI risk spanning multiple domains. These experts from across the AI industry, academia, governments, and civil society estimated the five most severe harms were likely to come from: AI having dangerous capabilities AI-enabled weapons and cyberattacks Competitive dynamics Power centralization The creation and dissemination of sophisticated false information Experts assessed each risk under two scenarios. Even if pragmatic mitigations are adopted over the next five years, experts judged that five of the 24 risk categories are more than 10% likely to cause catastrophic outcomes—defined as more than one million deaths, more than $100 billion in financial losses, or comparable harms. Without any mitigation, that number rises to 18. AI risks can reshape business and society even without causing catastrophic outcomes. Experts expressed concern that many risks, including false and misleading information, overreliance on AI systems, and fraud and scams, could reshape business and society even if they fall below catastrophic thresholds. The researchers convened a three-round Delphi study in late 2025 with 272 international AI risk experts that included AI researchers, policy advisors, technologists, and governance specialists to prioritize risks by judging their expected severity and likelihood, who is most vulnerable, and who should be responsible. The experts rated risk severity for different domains under two scenarios: “business as usual,” adhering to current trajectories without additional intervention, and “pragmatic mitigations,” reflecting the application of cost-effective, reasonable interventions. All expert contributions were collected anonymously, de-identified, and reported only in aggregate, limiting the potential for individual interests to influence results. No authors with competing interests were involved in the design or analysis of findings. The new study is part of the MIT AI Risk Initiative, a project that is building public knowledge infrastructure to help society to understand, prioritize, and manage risks from AI. This includes the AI Risk Repository, a living database of more than 1,700 AI risks, and tools such as the AI Incident Tracker, which connects risks to real-world AI harms, and the AI Governance Map, which analyzes risk coverage across laws, standards, policies, and other governance documents. Together, these resources support more informed, coordinated, and evidence-based AI risk management across the AI ecosystem. In the next phase of this research, the research team will analyze public documents from influential AI developers and large companies to examine whether organizations are responding to AI risks — and whether their responses are proportionate to the level of concern experts have expressed.