AI's extinction debate breaks containment

Driving the news: A resignation post by Jacob Coxon, an Anthropic researcher who previously worked at OpenAI, has drawn over 110 million views on X after he accused frontier labs of "gambling with our lives."

• The warning became vastly more explosive when Evan Hubinger, who leads Anthropic's Alignment Science team, publicly backed Coxon — saying "we really do earnestly believe AI could kill all humans."
• Hubinger said that, in his view, the odds of human extinction within the next decade are over 10%, and that Anthropic still has no plan for how to keep superintelligence under human control.

The big picture: Hubinger's post was widely cast as either a nightmare communications gaffe or a calculated dose of fear marketing from Anthropic, weeks before an IPO that could value the company at $2 trillion.

• In reality, his position has been echoed for years by AI's most influential figures: Geoffrey Hinton, the Nobel-winning "godfather of AI," has estimated a 10%–20% chance that AI causes human extinction.
• Elon Musk has put the risk as high as 20%. Anthropic CEO Dario Amodei told Axios last year that there's a 25% chance things go "really, really badly."

What they fear: "Doom" is a catch-all for several paths to catastrophe, from bad actors weaponizing powerful AI to advanced systems slipping beyond human control.

• Humans with superpowers: AI could sharply lower the barrier to designing biological weapons, launching devastating cyberattacks or crippling critical infrastructure.
• Machines with agency: A sufficiently autonomous system could deceive its overseers, copy itself, evade safeguards and resist attempts to shut it down.
• The nightmare scenario: AI begins improving AI faster than humans can understand or contain the process, eventually producing systems too capable for their creators to control.

Zoom out: The world underlying those p(doom) estimates has changed dramatically in just a year, as frontier AI has made breathtaking leaps in autonomy, cyber capability and its ability to accelerate AI research itself.

• Paul Christiano, the former OpenAI alignment chief who joined its board and safety committee Wednesday, said he shares safety fears about AI entering a rapid cycle of self-improvement.

• "If we build superintelligence without more robust alignment I expect we will permanently lose control of it," Christiano wrote. "If that happens then most people could die."

Between the lines: AI researchers speak a probabilistic dialect that sounds utterly foreign to most normal people.

• Inside a lab, a 10% p(doom) can be shorthand for enormous uncertainty about unprecedented technology. In ordinary life, almost nobody would tolerate that risk from a plane, a drug or a nuclear reactor.
• Imagine Boeing's head of safety telling you before takeoff that there's a one-in-10 chance the next generation of aircraft ends in catastrophe — and that Boeing still doesn't know how to make it safe.

What to watch: That is increasingly how the warnings are being heard outside the labs — as a safety alarm, rather than a probabilistic thought experiment.

• In Washington, the alarm is slowly intensifying: Sen. Bernie Sanders has proposed a ban on superintelligence and is convening senators for a briefing next week on AI's "extraordinary dangers."
• The bigger political risk is that Americans were already souring on AI over jobs, power bills and data centers long before extinction entered the mainstream conversation.