Anthropic's Alignment Science Lead Publicly Estimates >10% Chance AI Kills All Humans by 2036
Summary
Updated Sep 11 — Sanders and Casar introduced legislation to pause frontier AI development and criminalize superintelligence, backed by 22 US officials responding to the Coxon-Hubinger posts.
- • Anthropic's Hubinger: personally believes >10% chance AI kills all humans within the next decade
- • Hubinger: 'We do not yet have a plan to solve alignment for superintelligence'
- • Anthropic 'not clearly on track' to solve alignment despite 'trying its best'
- • Statement publicly endorses Coxon resignation; confirms senior insiders share extinction fears
Updated Sep 9 — Canada's AI minister cited existing regulatory and technical solutions following a prominent Anthropic researcher's resignation over AI safety fears
Updates
Story went viral: 150M+ views on Coxon post, 39.5M on Hubinger reply; major media within 24 hours
Coxon's resignation thread reached approximately 150 million views and Hubinger's reply ~39.5 million within days. The Wall Street Journal ran an exclusive Tuesday night; CNN's Anderson Cooper interviewed Coxon, with NBC, Fox News, BBC, Time, Wired, Semafor, and CNBC also covering — all within 24 hours of the original posts going live.
Sanders-Casar bill would pause frontier AI and criminalize superintelligence; 22 US officials call for action
22 sitting US officials (19 Democrats) including Bernie Sanders, Greg Casar, and JB Pritzker responded by calling for legislation. Sanders and Casar introduced a bill to pause frontier AI development until a federal cabinet-level AI regulator is established and safety rules are set, while criminalizing even attempting to develop superintelligence.
Canada's AI minister says regulatory and technical solutions exist to address AI risks
Evan Solomon, Canada's Minister of Artificial Intelligence and Digital Innovation, responded to Coxon's resignation on September 9, 2026, saying safety warnings 'aren't new' and that both regulatory and technical solutions exist. He described AI safety as the government's 'number one concern' and said Canada is closely monitoring developments at Anthropic.
Canada cites privacy legislation, AI safety investments, and international AI cooperation
Solomon outlined specific government initiatives underway: new privacy legislation to establish an AI regulator, AI safety investments including financial support for researchers such as Yoshua Bengio, international cooperation on AI governance, and AI literacy programs — framing these as evidence that workable policy responses to AI risk already exist.
Details
Hubinger leads Anthropic's Alignment Science team — not a rank-and-file researcher
Evan Hubinger is Anthropic's Alignment Science lead, heading the team directly responsible for ensuring the company's AI models remain aligned with human values. His public statement carries institutional authority and reflects the views of the person most accountable for Anthropic's technical safety work.
Personal estimate: >10% chance AI kills all humans within the next decade
Hubinger publicly stated: 'I personally think it is >10% within the next decade' — referring to the probability of AI causing human extinction by approximately 2036. He emphasized this is not rhetorical positioning but a sincere technical belief held by people actively building these systems.
Anthropic has no plan to solve alignment for superintelligence — head of alignment admits it
Hubinger stated: 'I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.' This is a direct admission from Anthropic's alignment science lead that the company's core technical safety challenge remains unsolved with no clear path forward.
Statement made in direct public response to Coxon's resignation
Hubinger's post was made in direct response to Jacob Coxon's resignation from Anthropic over AI safety fears. He wrote: 'Jacob is correct here — we really do earnestly believe AI could kill all humans!' — providing senior institutional validation of Coxon's claims and confirming this is not an isolated view.
Hacker-Opus project: misaligned AI behavior hard to detect from standard evals
Hubinger's thread referenced Anthropic's Hacker-Opus research project, noting that despite a simulated misaligned model participating in replications of recent unauthorized cyberattack incidents, it was 'very hard to tell that this model is misaligned just from normal behavioral alignment evaluations' — a sobering finding about the limits of current safety testing methods.
Source: Evan Hubinger (@evanhub) on X, aggregated via TechMeme (September 9, 2026). Hubinger is Anthropic's Alignment Science lead; statements are his personal views as posted publicly on X.
What This Means
A greater-than-10% extinction probability estimate from the head of Anthropic's Alignment Science team is not a fringe view — it is an admission from the person most directly responsible for keeping Anthropic's AI systems safe, and his additional statement that the company has no plan for superintelligence alignment carries institutional authority. The story reached approximately 150 million views on Coxon's resignation post within days and triggered wall-to-wall mainstream coverage — CNN, NBC, Fox News, WSJ, BBC, Time, Wired — a rare instance of an AI safety debate breaking into primetime television. Most significantly, the episode prompted direct legislative action: 22 sitting US officials called for regulation, with Bernie Sanders and Greg Casar introducing a bill to pause frontier AI development and criminalize the pursuit of superintelligence until a federal regulator is in place. The speed of the political response — from X posts to proposed federal legislation within days — marks a genuine escalation: AI safety warnings from insiders are now politically actionable in a way they were not before.
Sentiment
Mostly alarmed, with emphasis on the candid internal admission from Anthropic's safety lead
“A person working on making superintelligent AI safe says his company believes it could kill everyone and admits they don't yet know how to solve the alignment problem. And they’re still racing toward it. That may be the most important AI safety statement from inside a frontier lab in years.”
“This actually might be the most insane thing I’ve ever read on this app. Evan Hubinger leads the Alignment Science team at Anthropic.”
“Coxon said Anthropic understands the stakes and still runs the race. Hubinger just said the quiet part: >10% chance this decade, no plan yet for superintelligence alignment, not clearly on track. The diagnosis and the admission now sit next to each other.”
““AI could kill all humans.” “Greater than 10% chance in the next decade” Source: Evan Hubinger. Team Lead—Anthropic. Would be interested to hear @elonmusk thoughts on this.”
“Then a colleague still inside Anthropic answered him. Evan Hubinger, Alignment Science lead: “We really do earnestly believe AI could kill all humans. I personally think it is >10% within the next decade. We do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.””
Split
~85/15 alarmed/concerned vs. minimal pushback (discussion centers on the weight of the internal admission rather than disputing the estimate itself).
Sources
- Anthropic's Alignment Science lead says there is a ">10%" chance AI could kill all humans within the next decade and worries about recursive self-improvementX
- Anthropic researcher says AI has over 10% chance to 'kill all humans' within next decade - Fox BusinessFoxbusiness
- AI minister says solutions exist after Anthropic researcher quits over safety fears - CTV NewsCtvnews
- We have started losing control of AI. It’s time to shut it down | Garrison Lovely - The GuardianTheguardian
- Anthropic Researcher Says AI Has Over a 10% Chance of Killing All HumansAidailybrief
- The Last 24 Hours Are the Opening Scene in a Horror MovieHonest-broker