Ask "is AI dangerous?" and you'll get an answer shaped entirely by who you asked. A lab CEO might say the risk is low but worth taking seriously. A safety researcher might put the odds far higher. A parent might not be thinking about extinction risk at all, just about what their teenager is talking to at 2am.
Researchers just tried something more useful than picking one of those voices: they asked 272 of them at once, systematically, and compared notes.
The Study, in Plain Terms
MIT FutureTech and the University of Queensland surveyed 272 AI experts across 37 countries, spanning AI safety researchers, corporate risk officers, professors from MIT, Harvard, Oxford, Stanford, and Tsinghua, and government policymakers. Using a structured method called the Delphi process, they asked each expert to assess 24 distinct categories of AI risk.
Under current practices, with no additional safeguards, experts judged 18 of the 24 risk categories carry at least a 10% probability of a catastrophic outcome by 2030
"Catastrophic" was defined concretely: more than one million deaths, more than $100 billion in financial losses, or comparable civilization-scale harm, such as the collapse of democratic norms
The five risks experts rated as most severe: dangerous capabilities, competitive dynamics between labs and states, weapons and cyberattacks, power centralization, and false or misleading information
Even under "pragmatic, cost-effective" mitigations, five risk categories stayed above the 10% threshold: dangerous capabilities, weapons and cyberattacks, environmental harm, inequality and unemployment, and power centralization
Who Gets Hurt, and Who's Actually in Charge
One of the study's sharpest findings isn't about any single risk. It's about a mismatch.
Experts said the general public and everyday AI users bear the most vulnerability to these risks. But they assigned primary responsibility for addressing them to a different group entirely: general-purpose AI developers and governance actors like regulators and standards bodies.
"I cannot overstate how much more responsible the 'upstream' actors are for limiting these issues," one expert told researchers. "I see a direct analogy to social media. Yes, individuals are responsible for sharing misinfo. But the platforms should bear the brunt of our concern about the issue."
The researchers note this isn't unusual on its own; the public is similarly vulnerable to aviation failures or pharmaceutical side effects, while engineers and regulators carry the responsibility for prevention. What's different, they argue, is that AI mostly lacks the mandatory standards, enforcement, and liability regimes that bridge that gap in those other industries.
| Risk Domain | What Experts Said About It |
|---|---|
| Dangerous capabilities | Emerges naturally as systems scale, making it hard to predict or fully control even with safeguards |
| Weapons & cyberattacks | Seen as a "permanent fixture," from both state and non-state actors |
| Power centralization | Called the "most stubbornly persistent risk," since the entities building AI are often best placed to capture its benefits |
| Inequality & unemployment | Automation-driven job losses concentrated in certain sectors, deepening existing inequality |
| Environmental harm | Data-center energy demand and hardware footprint flagged as an ongoing resource pressure |
Where the Real Disagreement Lives
Ask about extinction-level risk specifically, rather than the broader catastrophic-harm categories above, and the expert consensus dissolves.
Estimates of what researchers sometimes call "P(doom)," the probability AI causes human extinction or something comparably severe, range from under 1% among some prominent industry figures to over 20% among certain AI safety researchers. The spread shows up even inside the same companies: Anthropic CEO Dario Amodei has put his own estimate in the 10-25% range, while OpenAI's Sam Altman has described his as low but non-zero.
Nobel-adjacent AI pioneer Geoffrey Hinton has said AI could prove as transformative as the industrial revolution, and potentially more dangerous. Yoshua Bengio, who chaired the 2026 International AI Safety Report, has pointed to a different kind of surprise altogether: harms nobody forecast at all. "One year ago, nobody would have thought that we would see the wave of psychological issues that have come from people interacting with AI systems and becoming emotionally attached," he said, citing cases involving children and adolescents that "should be avoided."
Is AI Dangerous: FAQ
There's no single yes-or-no answer experts agree on. A 2026 survey of 272 international AI experts found they judged 18 of 24 risk categories to carry at least a 10% probability of catastrophic outcomes by 2030 under current practices, but experts disagree sharply on the probability of extinction-level outcomes specifically, with estimates ranging from under 1% to over 20% even among senior figures at the same AI labs.
Researchers surveyed 272 AI experts across 37 countries using the Delphi method, asking them to assess 24 AI risk categories. They found the five risks with the most severe expected harm were dangerous capabilities, competitive dynamics between AI developers and states, weapons and cyberattacks, power centralization, and the spread of false information.
The study found a mismatch: the general public and everyday AI users bear the most vulnerability to AI risks, while experts assigned primary responsibility for addressing those risks to general-purpose AI developers and governance actors like governments and regulators, not to the people actually exposed to the harm.
P(doom) is shorthand some researchers use for their estimated probability of AI causing human extinction or a comparable catastrophe. Estimates vary enormously, from under 1% among some AI lab leaders to over 20% among certain safety researchers, and even differ sharply between co-founders of the same company, reflecting genuine, unresolved scientific and philosophical disagreement rather than a settled consensus.
There's broader agreement on already-observed harms, such as AI-enabled fraud, deepfakes, discriminatory outcomes, and cases of unhealthy emotional dependence on chatbots, some involving minors, which Turing Award winner Yoshua Bengio has cited as risks nobody anticipated a year ago. There's much less agreement on longer-horizon, larger-scale risks like an AI system pursuing goals in conflict with human oversight.
The MIT/UQ study found that even with what it called pragmatic, cost-effective mitigations, five risk domains still carried more than a 10% probability of catastrophic outcomes: dangerous capabilities, weapons and cyberattacks, environmental harm, inequality and unemployment, and power centralization. The researchers argue this level of risk would trigger mandatory action in most other safety-critical industries, and that voluntary developer action alone is unlikely to be sufficient.
Jans Bock-Schroeder
Publisher & Founder of AI Angst
Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.
Sources and Citations
This article is based on the following sources:
-
MIT AI Risk Initiative — "Priority AI Risks" (airisk.mit.edu, 2026)
Primary source for the full study findings, expert quotes, and risk-domain breakdowns.
https://airisk.mit.edu/priorities -
MIT Sloan — "These are the most urgent AI risks, according to 272 experts" (July 20, 2026)
Source confirming top-line figures and framing of the study's release.
https://mitsloan.mit.edu/ideas-made-to-matter/these-are-most-urgent-ai-risks-according-to-272-experts -
University of Queensland News — "Global experts assess risk of AI catastrophes" (June 5, 2026)
Source for study background and co-lead researcher Michael Noetel's comments.
https://news.uq.edu.au/2026-06-global-experts-assess-risk-ai-catastrophes -
Al Jazeera — "Why are experts sounding the alarm on AI risks?" (February 15, 2026)
Source for Yoshua Bengio's comments on unanticipated psychological harms.
https://www.aljazeera.com/news/2026/2/15/why-are-experts-sounding-the-alarm-on-ai-risks -
Calcuja — "P(doom) Survey 2026: AI Extinction Risk Estimates" (June 2026)
Source for aggregated existential-risk probability estimates across researchers and lab leaders.
https://calcuja.com/research/ai-risk-survey-2026/
Published: July 23, 2026. Sources verified at time of publication. All external links open in a new tab. This piece covers a topic where genuine expert disagreement exists; it aims to represent the range of credible views rather than endorse one.
