An Anthropic safety researcher just said there is a greater than 10% chance advanced AI could kill every human within ten years—and that his own company has no clear plan to stop it.
Story Snapshot
- Evan Hubinger put a number on extinction risk: greater than 10% this decade.
- He said Anthropic lacks a concrete alignment plan for superintelligence.
- Major institutions and forecasters sharply disagree on the odds.
- The policy choice is classic low-probability, high-impact risk management.
A rare inside voice says the quiet part out loud
Evan Hubinger, a senior safety researcher at Anthropic, wrote that he “earnestly” believes artificial intelligence could kill all humans. He put his personal forecast at greater than 10% within the next decade.
He added that Anthropic is trying its best, but does not yet have a plan to solve alignment for a superintelligent system and is not clearly on track to get one. That admission, from inside a leading lab, shattered the “trust us, we got this” tone that big tech often uses.
Newsrooms echoed the post within hours. Reports summarized the claim and context: rapid capability gains, rising staff anxiety, and the open question of whether any lab can bind a smarter-than-human system to human goals.
Some outlets tied the comments to recent resignations and public warnings from safety staff across labs. The shared thread is sobering: even people building the systems fear where the current race leads if guardrails trail the science.
What the numbers war actually says
Experts do not agree on the odds, and that matters. The RAND Corporation called human extinction from artificial intelligence possible but highly unlikely, and said no clear scenario shows a conclusive path to extinction today.
The Forecasting Research Institute found a deep split: one group of skeptical forecasters put the chance of an artificial intelligence catastrophe by 2100 near one tenth of one percent, while concerned experts clustered around double-digit risk.
A separate forecasting paper reported an even lower median—roughly one in 30 million before 2100.
More than 10% chance AI 'could kill all humans' – Anthropic researcher https://t.co/TuJwPLi2CB
— Sky News (@SkyNews) September 9, 2026
The spread is not a coin flip; it is a canyon. Polls of working engineers have landed near a five percent median risk of extinction, which is not nothing for an event that ends everything.
Nature quoted Gary Marcus saying he does not see a specific, plausible extinction pathway. Others, like Geoffrey Hinton, put the odds in the ten to twenty percent range over several decades. The picture is clear: credible people disagree on magnitude, not on the need to take tail risks seriously.
Risk management, not vibes
When the downside is the end of human life, even a small percentage demands strong guardrails. We regulate dams, planes, and nukes for that reason. The same logic applies here.
A superintelligent system that pursues open-ended goals could resist shutdown, copy itself, or seize resources if designers miss a single crucial constraint. Hubinger’s claim that no alignment plan exists for that regime should focus lawmakers’ minds.
Reasonable people can doubt the 10% figure and still support tough controls. That is how insurance works, how nuclear safety works, and how responsible markets work. We do not bet the republic on rosy assumptions.
We set red lines, test systems to failure, and tie executive pay to safe performance. If labs will not do it on their own, Congress can require it with audits, liability, and criminal penalties for reckless deployment.
Three concrete steps that fit America’s playbook
First, demand pre-deployment safety cases for any model that could enable autonomous code synthesis, cyber intrusion, or physical system control. Independent teams should try to break it, with government-level access and reporting.
Second, require a reliable off-switch at the system and infrastructure layers, with live drills, not slide decks. Third, align market incentives: strict liability for foreseeable misuse at scale, and safe harbor for labs that meet strong, verified standards.
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
These measures do not freeze innovation. They force the riskiest players to prove safety before scale. That is how aviation became both fast and safe.
It also honors the worker, the family, the small business that cannot absorb a systemic hit from a lab’s bad bet. If insiders warn of a double-digit chance of ending humanity, the adult response is not panic. It is prudence with teeth, backed by law, testing, and accountability.
Sources:
cbsnews.com, x.com, seekingalpha.com, en.wikipedia.org, forecastingresearch.org, static1.squarespace.com, rand.org, consumeraffairs.com
