Automations

MIT study maps 95 AI risk paths; 14% of leaders run AI at scale

A review co-authored by MIT CSAIL maps 95 ways AI could cause catastrophe. Our benchmark shows 14% of leaders already run AI in daily operations or at scale.

14% of business leaders say AI already runs in daily operations or at scale: exploring 35%, experimenting 28%, piloting 22%, operating 8%, scaling 6%, from The AI Report benchmark of 15,414 leaders
Editorial Team
November 13, 2026

Most published scenarios of AI catastrophe describe AI seizing control, and a new review finds that some other routes are barely mapped. The systematic review, co-authored by researchers at MIT CSAIL, the Existential Risk Observatory and the Future of Life Institute, collected 95 published scenarios for how AI could cause an existential catastrophe and sorted them into eight families. One of them describes businesses and institutions handing decisions to AI step by step. Our benchmark shows where companies are on that road: 14% of business leaders say AI already runs in their daily operations or is scaling.

Key takeaways

  • Rogue AI dominates the literature. 41 of the 95 scenarios (43%) describe a single AI taking over, according to the review.
  • A slow handover comes second. 18 scenarios describe people gradually handing real decision-making authority to AI until they cannot take it back.
  • Some routes are barely studied. Only 3 scenarios cover automated warfare, 4 cover AI systems coordinating a takeover and 1 covers permanent economic stratification.
  • Handover in business is still early. 14% of 15,414 business leaders say AI runs in daily operations or is scaling, rising to 22% at companies with more than 1,000 employees.
  • Risk rarely tops the list of AI problems. 3% of leaders name compliance as their biggest AI blocker and 6% name unreliable output, against 31% for choosing tools and 30% for the skills gap.

What did the MIT-backed review find?

The authors screened 3,184 documents and found 58 that work out at least one AI existential risk scenario in full, giving 95 scenarios dating back to 1993, according to the paper, "Existential Risk Scenarios from AI: A Systematic Review". Its 20 authors come from 17 institutions, including MIT FutureTech, the Existential Risk Observatory and the Future of Life Institute. They sorted the scenarios into eight families in three groups:

  • AI takes control (56 scenarios). A single AI takes over (41), many separate AI systems each gain influence beyond correction (11), or AI systems coordinate to do so (4).
  • Humans hand over control (18). In incremental handover, authority is delegated to AI systems that are not seeking power, until it cannot be taken back.
  • Power concentrates in a few hands (11). An actor uses AI to entrench its own position against any challenge (10), or AI replaces human labor until power sits with those who own everything else needed to produce (1).
  • AI multiplies destructive power (10). AI helps create a catastrophic invention (7) or drives automated warfare (3).

The authors stress that the counts show where researchers have done the work, not which outcomes are most likely. Their point is that defenses differ by route: work on aligning AI with human goals may reduce the risk of takeover, but might not address the other five families, the paper says. Its conclusion is blunt: "A strategy to prevent an existential catastrophe from AI has to address every scenario in this work." The paper, a free dataset and an interactive scenario explorer are available at aixrisks.com.

What is incremental handover?

Incremental handover is the scenario in which people keep delegating real decisions to AI until they lose the ability to take them back. In the review's description, authority over economic, cultural and institutional work moves to AI one step at a time, and each step makes it costlier to keep humans involved. As a task is handed over, the training, institutional knowledge and career paths that kept people able to do it fall away, so relying on AI becomes necessary rather than merely convenient. No AI has to seek power for this to happen, which is why the authors treat it as its own family.

How many businesses have handed real work to AI?

Still a minority. Of 15,414 business leaders who told us their AI stage, 35% are exploring, 28% are experimenting and 22% are piloting. 8% say AI runs in their daily operations and 6% say it is scaling across the business, 14% in all. Size matters: 22% of leaders at companies with more than 1,000 employees are operating or scaling AI, against 8% of solo operators.

How does your company compare? Take our 2-minute AI benchmark, the same nine questions 15,000+ business leaders have answered, and get The AI Report free every weekday.

What worries business leaders about AI?

Getting it to work, far more than controlling it. Asked for their single biggest AI blocker, 31% of 15,844 business leaders name choosing the right tools, 30% the skills gap and 15% a lack of time. Only 6% name unreliable output and 3% name compliance. Among the 2,211 leaders who already run AI in daily operations or at scale, the share naming compliance nearly doubles to 6%, and 7% name unreliable output. Leaders who already rely on AI name control problems more often.

Bar chart: share of business leaders naming each as their biggest AI blocker. Choosing the right tools 31%, skills gap 30%, no time 15%, unreliable output 6%, budget 6%, moving too fast 5%, compliance 3%, team resistance 2%, it did not stick 1%.
Compliance ranks near the bottom of business leaders' AI blockers.

For the security side of the same question, read our piece on AI model vulnerabilities and cyber risk.

Who should oversee AI decisions?

People, for now. Cillian Kieran, CEO and founder of Ethyca, on The AI Report podcast: "Do I feel personally comfortable with those machines making all of the high risk decisions they are likely to make in the future alone without any kind of oversight? I'm not super comfortable with that yet."

Kieran's case for treating AI governance as an engineering problem is in our write-up, Who controls your AI agents?

What should business leaders do now?

  • Map where AI decides and where it assists. List the processes where AI output goes straight into action without a person signing off.
  • Keep the skills you hand over. When AI takes on a task, keep a few people trained to do it by hand, so the company can step back in.
  • Name an owner for AI oversight. Someone should be accountable for reviewing what automated systems decide, not just whether they run.
  • Log what agents do. A record of each automated action makes mistakes visible before they compound.
  • Start small and reversible. Our guide to choosing processes for AI automation helps pick tasks where a wrong call is easy to undo.

See where your company stands

Compare your company with 15,000+ business leaders in the AI Adoption Index, then take the 2-minute benchmark yourself. It is free, and so is The AI Report, read by 400,000+ business leaders every weekday.

About this data

The figures come from The AI Report's AI benchmark, answered by 15,844 business leaders between July 20 and November 13, 2026, counting one answer per person. The stage figures come from the 15,414 who chose one of five stages (exploring, experimenting, piloting, operating, scaling); "operating or scaling" combines the last two. The blocker figures come from the question asking for each leader's single biggest AI blocker, answered by all 15,844, and the figures for leaders already operating or scaling come from 2,211 of them. Company size comes from the same answers. The quote from Cillian Kieran comes from his September 2026 episode of The AI Report podcast. Readers chose to subscribe to an AI newsletter, so they are likely further along than average. We never show groups under 300 people. Scenario counts come from the review itself, whose authors note that one reviewer coded the scenarios and that the counts are indicative rather than precise.

Cite this: The AI Report, "MIT study maps 95 AI risk paths; 14% of leaders run AI at scale," November 17, 2026, theaireport.ai.

Join the Newsletter
Inchide fereastra