MIT survey of 272 experts ranks AI's top risks to 2030

A new expert survey out of MIT tries to do something the AI risk debate mostly avoids, which is to put probabilities on the worst outcomes. Researchers at MIT FutureTech and the University of Queensland used the Delphi method to poll 272 international AI experts on 24 risks over a 2025 to 2030 window, MIT Sloan reported, and scored each risk under two scenarios: business as usual, and pragmatic mitigation, in which organizations and governments make cost-effective efforts to reduce AI risks.

Even in the pragmatic mitigation scenario, five domains cleared a 10 percent probability of catastrophic outcomes. AI systems possessing dangerous capabilities, AI-enabled weapons and cyberattacks, and environmental harm each landed at 12 percent. Inequality and unemployment came in at 11 percent, as did power centralization and unfair distribution of AI's benefits. Peter Slattery, one of the lead researchers, noted that coding and hacking are some of the areas where they are seeing the fastest growth in AI capability, which helps explain why the weapons-and-cyberattacks bucket sits at the top.

The sector picture in the piece is narrower than the risk list. Three areas showed up as most exposed: information, where misinformation, privacy loss, and erosion of trust cluster; national security, where cyberattacks, weapons development, and surveillance dominate; and finance, where AI could amplify fraud, market manipulation, and failures in systems that have broader economic effects. The study also flags a mismatch worth pausing on. The experts said developers and governance actors bear primary responsibility for addressing AI risks, while users and the stakeholders those systems affect are the ones most vulnerable to them.

The honest caveat is that these are expert judgments aggregated through a structured process, not empirical base rates, and the article does not say much about the disciplinary makeup of the 272 respondents or how tightly the panel agreed. Take the 11 to 12 percent figures as 'a serious minority of qualified people think this is where things go badly,' not a calibrated forecast.

For anyone building AI governance inside a company, the more useful takeaway is the AI Risk Repository the same group maintains, a continuously updated database of more than 1,600 AI risks organized by cause and domain that policymakers, technologists, and organizations are already using. Rather than reinventing a taxonomy, plug into that one and let the survey tell you which five buckets to weight heaviest.