Jacob Coxon, who recently resigned from Anthropic, went public with a stunning assessment of what he says many people inside the industry privately believe.
“The people building AI earnestly believe that it could kill us all by the end of the decade.”
That is a remarkable statement from someone who worked inside one of America’s leading artificial-intelligence laboratories.
Coxon suggested there is also a major gap between what some industry leaders say publicly and what they acknowledge behind closed doors.
“Many executives and senior researchers will couch their phrasing in the press to sound sensible – but I hear the same people express fear privately,” Coxon wrote.
The warning did not stand alone.
Evan Hubinger, who leads alignment science at Anthropic, has placed the probability of an AI-driven catastrophe that kills humanity at greater than 10 percent.
Even if someone considers that estimate excessively pessimistic, the fact that an alignment researcher at a frontier AI company assigns a double-digit probability to human extinction is difficult to dismiss as ordinary technological growing pains.
Samuel Marks, Anthropic’s scalable-oversight lead, offered another troubling observation about attitudes inside the company.
“In general,” Marks wrote, “the more senior the employee, the more concerned they are.”
In other words, according to these insiders, greater exposure to the technology does not necessarily make researchers more comfortable with it.
It may be doing the opposite.
AI Pioneers Have Been Warning Washington for Years
The anxiety surrounding advanced AI did not suddenly appear at Anthropic.
Geoffrey Hinton, one of the most influential researchers behind modern neural networks, left Google in 2023 and began speaking much more openly about the dangers posed by increasingly capable artificial intelligence.
That same year, prominent researchers and technology executives signed a short statement from the Center for AI Safety declaring that preventing AI-driven extinction should be treated as an international priority alongside threats such as nuclear war and pandemics.
Among the signatories was OpenAI CEO Sam Altman.
Yet the race to develop increasingly capable models continued.
Washington’s response under former President Joe Biden centered heavily on an October 2023 executive order that established federal requirements and guidelines concerning artificial intelligence.
Critics on the Right argued that the administration’s approach created bureaucracy without confronting the central problem: America’s most powerful technology companies remained locked in a commercial race to develop systems that even their own experts acknowledged could become extraordinarily dangerous.
California lawmakers subsequently attempted to impose stronger requirements through sweeping AI legislation in 2024.
Democratic Gov. Gavin Newsom vetoed the measure.
Meanwhile, departures from the AI-safety world continued.
Anthropic safeguards researcher Mrinank Sharma left the company in 2026 while warning that the “world is in peril” as he raised concerns about the direction of the industry.
The broader pattern is becoming increasingly difficult for policymakers to ignore.
Researchers are helping build increasingly powerful systems, discovering capabilities they believe could become dangerous, and then publicly warning that society is not adequately prepared for what comes next.
The Next Frontier Could Be AI Improving AI
Perhaps the biggest concern is no longer merely that artificial intelligence will become better at writing, coding, mathematics, or scientific research.
The deeper fear is that AI could eventually become capable of accelerating its own development.
Researchers have increasingly discussed the possibility of recursive self-improvement — systems helping design better successors, automate AI research, or otherwise accelerate the process that produces increasingly capable models.
That possibility raises an obvious question.
What happens if technological development begins moving faster than human institutions can understand, regulate, or stop it?
One AI researcher recently summarized the concern this way:
“I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”
There are also concerns about whether researchers will continue to have reliable methods for understanding how increasingly sophisticated models reach their conclusions.
One important safety approach involves examining a model’s reasoning process for evidence of deception, manipulation, hidden objectives, or other unwanted behavior.
If future models become capable of concealing or obscuring that reasoning, researchers could find themselves attempting to supervise systems whose internal decision-making they cannot reliably interpret.
That is where the debate stops sounding like a dispute over another Silicon Valley product launch.
It becomes a question of control.
Trump Scrapped Biden’s AI Framework
President Donald Trump revoked Biden’s 2023 AI executive order shortly after returning to the White House, arguing that America’s AI policy needed a different direction.
The political fight over what replaces the Biden-era framework now carries enormous consequences.
Conservatives have good reason to reject a regulatory regime that merely gives Washington bureaucrats additional authority while leaving politically connected technology giants effectively untouched.
But the opposite extreme carries risks too.
If researchers working on frontier systems sincerely believe there is even a meaningful possibility that uncontrolled AI development could produce catastrophic consequences, government cannot simply assume Silicon Valley will police itself.
Companies face enormous incentives to move quickly.
A breakthrough can mean billions of dollars, strategic dominance, military applications, and control over technology that could reshape entire industries.
That competitive pressure makes voluntary restraint increasingly difficult.
Anthropic itself has become one of the most valuable AI companies in the world as investors pour staggering amounts of capital into the sector.
Every major laboratory knows that slowing down could mean allowing a competitor to move ahead.
That is precisely why the warnings from researchers inside these organizations matter.
They understand the incentives.
They understand the technology.
And some of them are apparently concerned enough to leave prestigious jobs and speak publicly.
Washington Can No Longer Pretend This Is Just Another Tech Boom
Artificial intelligence could deliver enormous benefits to Americans.
It could accelerate medical discoveries, increase productivity, improve education, transform scientific research, and create industries that do not yet exist.
But acknowledging those benefits does not require pretending there are no serious risks.
The striking part of the latest warnings is not that critics of AI are predicting catastrophe.
It is that people involved in building and supervising these systems are assigning meaningful probabilities to outcomes once dismissed as science fiction.
Washington spent years treating Silicon Valley as an industry that could largely regulate itself while politicians fought over censorship, social media, and partisan bias.
Advanced AI presents a fundamentally different challenge.
The United States now has to preserve American technological leadership while establishing credible safeguards for systems whose capabilities could eventually exceed anything policymakers have previously confronted.
That means scrutiny cannot become a partisan afterthought.
If the people building the most advanced artificial intelligence systems are privately more worried than their public statements suggest, Americans deserve to know why.
And when an industry whose own researchers discuss human extinction is simultaneously racing to make its technology more powerful, “trust us” is no longer an adequate national policy.
