MIT Study: 272 Experts Assess AI Risks, Model Makers Won't Bear the Costs

A team of researchers from MIT FutureTech and the University of Queensland has published the results of the largest systematic survey of expert opinion on AI risks to date. The conclusions are unambiguous: the developers of the most powerful AI models bear the greatest responsibility for the dangers these systems create, but it is users and society who will foot the bill when something goes wrong.

How the study was conducted

The authors used the Delphi method, in which experts assess issues over several rounds, seeing the anonymous ratings of other participants along the way and being able to revise their own positions. In this case, the study comprised three rounds carried out in late 2025 among 272 specialists from 37 countries, of whom 214 stayed through to the end of the process. Each participant rated only those of the 24 risk categories in which they declared genuine expertise, a design meant to increase the reliability of the results.

Experts assessed each category for both the likelihood of occurrence and the severity of its effects through a horizon extending to September 2030, and also indicated which sectors of the economy are most exposed and who bears responsibility for countering the threats.

Which risks are most concerning

Five categories received the highest average severity scores: dangerous AI capabilities and competitive pressure tied at 3.49 on a five-point scale, followed closely by AI-enabled weapons and cyberattacks (also 3.49), concentration of power (3.47), and the spread of false information (3.44).

In a scenario without additional safeguards, experts estimated the probability of catastrophe by 2030 at 21.5 percent for dangerous AI capabilities, 21 percent for weapons and cyberattacks, 18 percent for concentration of power, 16.6 percent for competitive pressure, and 12.8 percent for disinformation. As many as 18 of the 24 categories studied exceeded the 10 percent threshold for catastrophic outcomes.

Safeguards fall short

The researchers also examined how the ratings change once pragmatic risk-reducing measures, such as regulatory oversight or technical safety mechanisms, are factored in. Severity dropped across all categories, but five remained above the 10 percent threshold: dangerous model capabilities and weapons and cyberattacks at 12 percent each, environmental harm at 12 percent, inequality and unemployment at 11 percent, and concentration of power at 11 percent. In other words, even with realistic preventive measures in place, some threats remain in the high-risk zone.

Who's responsible, who pays

The study's most telling conclusion concerns the distribution of responsibility and costs. Model developers received a median score of 4-5 on the five-point responsibility scale in nearly every category, with governments and regulators rated similarly high. At the same time, the study finds that AI users and the general public, the groups with the least real capacity to control this risk, are the ones most exposed to its effects.

The authors stress that the competitive race for computing power, market share, and capital rewards speed of deployment at the expense of caution, and that the costs of eventual mistakes can easily be shifted onto third parties who had no say in the decisions behind a given system's development.

It is incredibly worrisome that experts are seeing a 10% probability of catastrophic outcomes across so many areas - Neil Thompson, director of MIT FutureTech, MIT Sloan

Which sectors are most exposed

According to the experts, the information and national security sector faces the greatest exposure to AI risks. Close behind are finance and insurance, where AI increases vulnerability to fraud, cyberattacks, and market manipulation, and healthcare, exposed to privacy violations, algorithmic discrimination, and excessive reliance on automated clinical decisions.

The study also covered the information technology and educational services sectors as areas particularly sensitive to the negative effects of AI deployment, though it was national security and information infrastructure that scored highest on threat level.

For Polish companies and institutions deploying AI systems, the study's findings carry practical weight, especially as work continues on national oversight of the AI Act and the supervisory commission has still not been appointed. If even international experts judge that responsibility is concentrated on the side of model providers while the costs fall on users, companies purchasing licenses for generative systems should verify safeguards themselves rather than assume that responsibility will automatically fall on the manufacturer.

The study's authors say they plan to continue developing the AI risk repository maintained by MIT FutureTech, which will be updated regularly as new data emerges on the capabilities and applications of future models.