AI Insiders Warn of 50% Human Extinction Risk

AI Researchers Break Silence on Superintelligence Risk

"The chance of human extinction is about a coin flip, in my view."

That assessment comes from Geoffrey Irving, a former employee of both OpenAI and Google DeepMind, in one of roughly a dozen interviews published by Palisade Research, a nonprofit that describes itself as studying the behavior of AI systems. Irving's estimate places the odds of humanity being wiped out by artificial intelligence at roughly 50 percent — a figure that lands like a thunderclap in a field where even single-digit risk estimates have historically been treated as provocative.

The interviews, reported by The Verge, draw on current and former employees at OpenAI, Google, and Anthropic, three of the most consequential AI laboratories on Earth. The through-line is uncomfortable: people who helped build the most capable systems in history are now publicly airing doubts about whether those systems can be controlled.

Irving's framing is blunt, but it is not an outlier in kind. It is an outlier in degree. For years, the loudest warnings about AI superintelligence risk came from philosophers and outside critics. Now the warnings are coming from inside the labs.

Who Are the AI Insiders Speaking Out?

Irving's résumé is the detail that gives his coin-flip estimate its weight. He worked not at one but at two of the world's leading AI developers — OpenAI and Google DeepMind — giving him direct exposure to the research cultures that produced the current generation of frontier models. When someone with that background says extinction odds sit near 50 percent, the claim carries a different texture than the same sentence from a commentator with no lab experience.

He is far from alone in the Palisade Research interview series. The collection includes current and former staff from OpenAI, Google, and Anthropic, spanning people still inside the industry and people who have left it. That mix matters. Current employees face professional constraints; former employees face accusations of sour grapes. The series contains both, which makes the pattern harder to dismiss as the complaint of any single camp.

Palisade Research itself is described as a nonprofit focused on studying AI systems. Its decision to publish a batch of candid safety interviews — rather than a single op-ed or a leaked memo — suggests a deliberate effort to document a broader shift in expert sentiment rather than amplify one dissenting voice.

What Does a 50% Extinction Risk Actually Mean?

Percentages about human extinction are easy to say and hard to interpret. So it helps to place Irving's number on a scale.

In 2023, a broad coalition of leading AI scientists and tech executives signed the Statement on AI Risk, which declared that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. That statement deliberately avoided attaching a number to the danger. Its signatories included many of the same kinds of researchers now giving interviews — a reminder that the community has treated existential risk as credible for years, even when it declined to quantify it.

The most-cited quantification comes from an expert survey conducted by AI Impacts in 2022. When researchers asked AI experts to estimate the probability of catastrophic outcomes from advanced AI, the median response was about 10 percent. That is a substantial figure on its own — roughly one-in-ten odds of a catastrophe — and it already exceeded what many outside observers considered reasonable.

Irving's 50 percent sits five times higher than that median. He is not revising a fringe view upward by a few points. He is staking out a position well beyond the mainstream of expert opinion as measured four years earlier.

Two caveats belong here. First, a 50 percent figure is a personal judgment, not a measurement; no one can run an experiment on human extinction. Second, the AI Impacts survey is now several years old, and the systems it asked about have been eclipsed by newer, more capable models. Whether expert medians have drifted upward since 2022 is precisely the question these interviews implicitly raise.

Why Are AI Researchers Speaking Out Now?

The Palisade Research series did not appear in a vacuum. It lands in the middle of a visible surge in AI whistleblower activity across 2025 and 2026.

Employees at major labs have grown more willing to attach their names to safety concerns — through open letters, testimony, and on-the-record interviews — even when doing so carries professional risk. The dynamic resembles earlier reckonings in other industries: insiders who helped build a powerful technology become the most credible witnesses to its hazards. Nuclear physicists played that role in the 1940s and 1950s. Tobacco researchers did in the 1990s. Now AI safety staffers are doing it in the 2020s.

Several forces appear to be pushing them forward. The capabilities of frontier models continue to advance faster than the public's ability to evaluate them. Governance efforts, while real, remain fragmented across jurisdictions. And the gap between what labs say privately about risk and what they say publicly has itself become a subject of internal tension — the kind of tension that eventually produces interviews like these.

There is also a simpler explanation. The people closest to the technology are the ones who understand what it can do. When that understanding produces alarm, silence becomes harder to justify.

The State of AI Safety at Major Labs

OpenAI, Google, and Anthropic all maintain dedicated safety teams, and all three have publicly committed to responsible development. The Palisade interviews complicate that picture without contradicting it. A lab can employ serious safety researchers and still contain staff who believe the overall risk trajectory is worse than official communications suggest.

The presence of current employees in the series is the detail worth watching. Current staff speak under constraints that former staff do not — they can be disciplined, reassigned, or shown the door. When they nonetheless participate in interviews about extinction risk, it signals that internal concern has reached a level where career caution is no longer fully suppressing it.

For the labs, the reputational math is awkward. Their safety teams are a central part of their public story. If a meaningful share of safety personnel privately estimates catastrophic risk at rates far above the expert median, the gap between marketing and internal belief becomes a governance problem, not just a communications one.

What Happens Next: Implications for AI Policy and Development

A 50 percent estimate from a former OpenAI and DeepMind researcher will not settle any policy debate. It will, however, change the texture of one.

Lawmakers weighing AI legislation now have a new data point to cite — one that comes from inside the industry rather than from advocacy groups. Regulators deciding how aggressively to test frontier models can point to expert disagreement as evidence that the current pace of deployment outruns the current pace of oversight. And the labs themselves face a sharper version of a familiar question: if your own former employees put extinction odds near a coin flip, what exactly is your safety case for continuing at full speed?

The AI superintelligence risk debate has always suffered from a measurement problem. Nobody can observe the probability of human extinction, so the argument defaults to dueling intuitions. The value of the Palisade Research interviews is that they make those intuitions public and attributable. Irving says roughly 50 percent. The 2022 AI Impacts median said roughly 10 percent. The 2023 Statement on AI Risk said the danger is serious enough to rank alongside nuclear war, without assigning it a number.

Three data points, three different levels of alarm, all from people who know the technology intimately. That spread is itself the story. The experts closest to the work do not agree on how much danger humanity faces — only that the question is no longer hypothetical, and that the people who built these systems are increasingly unwilling to stay quiet about it.

Source: The Verge