Most AI Researchers Don't Think AI Will End Humanity

The headline claims about AI extinction have grown louder over the past two years, with public statements warning that powerful AI systems could pose a serious risk to humanity's survival. A closer look at how researchers actually working in the field rank that risk tells a different story.

A March 2026 survey by researchers at University College London asked roughly 4,000 AI researchers to identify the most concerning long-term risks of their field, as reported in Nature. Only 3 percent named existential or extinction risk as their primary worry. The far more common answers concerned misinformation, mass surveillance, labour market disruption and concentration of economic power.

The mechanisms don't hold up

That gap between public discourse and researcher consensus is worth sitting with. The most frequently cited extinction scenarios involve an advanced AI either pursuing goals that diverge from human survival, or being used by a malicious actor to build a weapon of mass destruction. A RAND Corporation analysis examined the most cited mechanisms, including engineered pandemics, nuclear escalation, and large-scale geoengineering, and concluded that human extinction would not be a plausible outcome under any of them, short of an actor deliberately and successfully pursuing it.

There are other reasons for scepticism. Gary Marcus, an AI researcher at New York University who has been a vocal critic of overhyped claims in the field, said in a recent interview that he could not point to "any specific scenario for AI-induced extinction that seems particularly plausible". A separate paper published this year argues that "artificial general intelligence" as currently described is not a coherent scientific concept, but a marketing-adjacent term that has slipped into policy debates without rigorous definition.

The harms that aren't speculative

None of this is the same as saying AI carries no risk. A model that fluently produces misinformation at scale, or a deployment that quietly reshapes hiring or lending decisions across an industry, can do significant damage without ever resembling a science-fiction superintelligence. The point researchers tend to make is that focusing public attention on extinction can crowd out the discussion of harms that are real, measurable, and already happening.

It is also worth noting where the loudest doom warnings tend to originate. Much of the rhetoric comes from inside the AI labs themselves, often paired with statements that those same labs are uniquely positioned to manage the risk. That structure should at least invite some scepticism about how the framing serves the speakers.

For the next generation of AI policy, the more useful questions are mundane and difficult: who is liable when an AI system causes harm, what disclosure should accompany an automated decision, and what training data may be used and on what terms. Those questions do not lend themselves to dramatic press releases. They are what the researchers closest to the work actually spend their time worrying about.