Could AI Really Wipe Out Humanity? What the Actual Experts Think

The honest answer is: nobody knows, and the people who study this professionally disagree with each other by an enormous margin. That's not a dodge. The range of expert opinion on this question is genuinely, measurably wide, and understanding why it's so wide tells you more than a confident yes or no ever could.

First, a distinction that matters more than anything else here

"AI" isn't one thing. The chatbot you might use to draft an email today is not the same category of system as the hypothetical future AI that shows up in extinction-risk discussions. When researchers debate existential risk, they're almost never talking about current systems suddenly turning dangerous. They're debating whether some future, far more capable system, one that can act autonomously, set its own subgoals, and operate with little human oversight, might end up pursuing objectives that are catastrophic for humanity, whether on purpose or as a side effect of poorly specified goals.

Keeping that distinction in mind is the single most useful thing you can do before reading anything else on this topic. A lot of confused, sensationalized coverage collapses "the AI in your phone" and "a hypothetical future superintelligence" into the same conversation. They're not the same conversation.

What the actual surveys show

AI Impacts, a research group that has surveyed AI researchers since 2016, asked nearly 2,800 AI researchers directly about this in their most recent published survey. The median researcher put the probability of AI causing human extinction or similarly severe, permanent disempowerment at around 5%. A related question, about humanity losing the ability to control advanced AI systems, got a median estimate of 10%.

Prediction markets tell a similar story. Metaculus, a forecasting platform where participants are scored on accuracy, has had a community estimate hovering around 10%, gradually climbing since 2020.

The skeptical case

This isn't a fringe position. It's held by people who build the systems in question for a living.

The concerned case

Yudkowsky and Nate Soares published a book on this in 2025, and a Harvard event built around it in early 2026 is worth mentioning, with an important caveat.

Why the range is this wide

The disagreement doesn't come down to some researchers having better information than others. It comes down to genuinely different models of how AI development unfolds. Researchers focused on near-term capabilities tend to reason from what current systems can and can't do, and extrapolate cautiously. Researchers focused on long-term safety tend to reason from first principles about what a sufficiently capable, goal-directed system would do by default if its objectives weren't specified with extreme care, regardless of when such a system might arrive.

Even within the community most worried about this, there's no agreement on timelines. One well-known forecasting effort, AI Futures Project, published individual forecasters' estimates for when advanced AI might arrive with 10th-to-90th percentile ranges spanning well over a century. When the people making the case for concern can't agree on a timeframe within a hundred years of each other, that tells you the underlying uncertainty is real, not a matter of some people simply being better informed.

The view that gets less attention

Between "essentially zero risk" and "near certain catastrophe" sits a much larger, less headline-friendly group: researchers who think the risk is real enough to take seriously and fund safety research into, without believing it's the dominant near-term concern. This is probably the most common position among people who work in AI policy and safety professionally, even though it doesn't produce a dramatic number or a viral headline. It shows up in survey data as the wide middle of the distribution, the space between LeCun's near-zero and Yudkowsky's near-certain, where most of the actual probability mass sits once you account for the whole field rather than its most quoted extremes.

Where this leaves an honest answer

Could AI cause human extinction? Serious people who study this for a living genuinely disagree, and the disagreement spans from "essentially impossible with any technology on a realistic horizon" to "likely if development continues unchecked." Both ends of that range are held by credentialed, thoughtful researchers, not cranks on either side.

What almost everyone agrees on, across that whole range, is that the current generation of AI, the tools available today, are not the thing under discussion when this question comes up. The debate is about what might eventually be built, not what already has been.

That's a less satisfying answer than a headline promises. It's also the accurate one.