An Anthropic safety researcher just put a double-digit number on the chance AI wipes out humanity within ten years—and admitted his lab has no clear fix.
Story Snapshot
- Evan Hubinger said the chance AI “could kill all humans” is over 10% this decade.
- He added Anthropic lacks a plan to align superintelligence and is not on track.
- Major outlets amplified the warning amid staff turmoil across top labs.
- Leading skeptics counter that extinction is possible but highly unlikely.
What Hubinger Said, Plain and Unfiltered
Evan Hubinger, a safety researcher at Anthropic, wrote that his real forecast is stark: “AI could kill all humans,” and he puts that risk above ten percent within the next decade.
He added that while Anthropic is trying, it does not yet have a plan to solve alignment for a superintelligent system and is not clearly on track to find one. This is not a leak. It is a direct, public statement, and it leaves little wiggle room for soft spin.
Newsrooms moved fast. The British Broadcasting Corporation summarized the claim and tied it to wider anxiety about how fast companies train systems that can write code, plan tasks, and even find ways around guardrails.
CBS News signal-boosted the tipping-point tone as part of a drumbeat of resignations and warnings from inside labs that were once closed shops. This is the part that rattles people: insiders who build the tools now worry the brakes are not strong enough.
More than 10% chance AI 'could kill all humans' – Anthropic researcher https://t.co/TuJwPLi2CB
— Sky News (@SkyNews) September 9, 2026
How Big Numbers Become Headlines
Forecast tournaments and expert surveys rarely agree on a single number. A study by the Forecasting Research Institute documented a deep split: a “concerned” group put the chance of an artificial intelligence existential catastrophe near one in four by 2100, while skeptics put it close to one in one thousand.
A separate forecasting paper found a median estimate for artificial intelligence-caused extinction before 2100 at one in thirty million, with superforecasters far more skeptical than domain experts. Big ranges tell you the truth is uncertain and the stakes are extreme.
Policymakers care less about consensus than about consequence. The Bulletin of the Atomic Scientists cited a poll of over two thousand working engineers and researchers with a median five percent extinction risk—low odds, but not lottery low—and argued that even skeptics must treat tail risks as serious because the downside is civilization itself.
That logic matches common sense: when failure is irreversible, you do not gamble with the family farm.
The Counter-Case: Unlikely, Unclear, Unpersuasive?
Researchers at the RAND Corporation pushed back. They wrote that no clear, concrete scenario shows artificial intelligence as a conclusive extinction threat today, and one senior scientist said that outcome is “highly unlikely,” though not impossible. That critique matters.
It challenges fear that runs ahead of evidence and reminds us that complex systems usually fail in small, not final, ways first. Skepticism keeps engineers honest and keeps policy focused on real failure modes.
The skeptic case, however, must also square with the people closest to the fire. When a leading safety voice at a frontier lab says the team lacks a plan to align superintelligence, that signals operational risk, not just theory.
A responsible response blends both views: demand clear scenarios and testable claims, while tightening controls and timelines until those claims survive contact with reality.
What Smart, Practical Guardrails Look Like Now
Congress, regulators, and labs can do five things without choking innovation. First, require pre-deployment testing for dangerous capabilities like autonomous code execution, cyber intrusion, and lab-grade bio assistance, with third-party audits.
Second, tie model scaling to safety milestones: prove your brake works before you add more horsepower. Third, mandate incident reporting so near-misses become shared lessons, not buried secrets.
Fourth, harden compute access with identity checks and real-time monitoring. Fifth, impose personal accountability for executives who sign off on inadequate safety proofs. These are normal tools in aviation and energy; they fit here, too.
This actually happened…This morning’s AI headline is almost impossible to process: an Anthropic researcher just resigned saying the people building AI genuinely believe it could “kill us all by the end of the decade.” Anthropic’s alignment lead says he puts the risk above 10%.…
— Edward Corona – The Options Oracle (@EdwardCoronaUSA) September 9, 2026
Markets also have a role. Insurers should price policies to reward provable safety margins and penalize sloppy deployment.
Investors should ask one blunt question in every boardroom: what stops this model from escalating its reach without a human in the loop? Families do not want white papers; they want a kill switch that works under stress. That is not anti-tech. That is pro-stewardship.
The Bottom Line: Numbers Are Debates, Actions Are Choices
Hubinger’s 10% plus is not a prophecy. It is a risk call from inside a shop that knows how fast the frontier runs. Skeptics argue extinction is implausible and poorly specified today, and they might be right. But leaders do not bet the nation on “might.”
The prudent path is simple: verify, then trust. Slow the race until the seatbelts latch, the airbags fire, and the walls do not cave when you hit them. If the lab cannot show that, the lab should not ship.
Sources:
cbsnews.com, x.com, rt.com, seekingalpha.com, en.wikipedia.org, forecastingresearch.org, static1.squarespace.com
