Doomsday Clock For AI?

A respected Anthropic researcher quit and warned self-improving AI could arrive within two years—arguing labs are “gambling with our lives.”

Story Snapshot

Anthropic’s Jacob Coxon resigned, warning labs are racing toward self-improving superintelligence.
Anthropic says we are not at recursive self-improvement yet, but it could come sooner than institutions can handle.
A current Anthropic alignment scientist cited over a 10% chance of human extinction within a decade, focused on future systems.
Anthropic reports its Claude system now helps write most of its own production code, speeding development loops.

What Coxon Said And Why It Matters

Jacob Coxon, a researcher who worked on pretraining at both OpenAI and Anthropic, resigned and said major labs are racing toward self-improving superintelligence. He argued the race dynamic makes restraint unlikely without action from outside the labs. He warned the shift could happen “next year or the year after,” and framed the push as “gambling with our lives”. His insider status raises the stakes, even if he did not publish a technical model to back the one- to two-year clock.

Coxon’s broad claim reached millions through cable and online news, which can amplify fear. But much of his core point echoes a known market problem: when firms fear falling behind, they move faster. That can push safety to the side. Many readers on the left and right see the same pattern in other areas: profits first, people second. Coxon’s exit channels that anger at institutions that seem to listen only when pressure builds.

What Anthropic Publicly Acknowledged

Anthropic published a detailed note on “recursive self-improvement,” which means an artificial intelligence system can fully design and develop its own successor. The company said we are not there yet and it is not inevitable. Still, it warned the threshold could arrive sooner than most institutions are ready for. It urged giving the world the option to slow or pause frontier development to let safety and oversight catch up.

Anthropic also reported that its internal data shows its Claude system is already speeding up its own development by writing most of the code merged into its production codebase. That is not full self-improvement, since humans still review, secure, and deploy. But it does show a tighter loop where models help build the next models. Faster loops can magnify both benefits and risks, which is why the firm called for stronger safeguards now.

How Big Is The Risk Right Now?

A current Anthropic alignment scientist, Evan Hubbinger, reportedly estimated more than a 10% chance of human extinction from artificial intelligence within a decade. He also said current models are low risk, and the main danger comes from future systems that could improve themselves fast. That framing supports concern about the future without claiming today’s models are already out of control. It adds expert weight to Coxon’s urgency claim while keeping the focus forward.

Former Anthropic and OpenAI researcher Jacob Coxon leaving Anthropic before his stock even vests just to warn people about superintelligence says a lot.

Even wilder when Anthropic's own safety lead replies agreeing that there's a >10% chance AI kills everyone.

— Rahul Dahre (@RahulDahre) September 10, 2026

Some reports cited recent artificial intelligence–enabled cyber incidents as warning signs. These examples suggest growing agent skills, but public summaries do not prove autonomous self-improvement is already happening. Coxon did not release a technical forecast to support his tight timeline either. That leaves a gap between stark warnings and confirmable evidence. The stronger, documented point is clear: capabilities and task autonomy are improving fast, and oversight is struggling to keep pace.

Why Voters Across The Spectrum Care

Speed without guardrails feels familiar. Many Americans think elites reap gains while the public eats the risk. With artificial intelligence, the downside could be very large, even if the odds are debated. People frustrated by waste, weak regulators, and self-dealing see another field where insiders race ahead and ask for forgiveness later. The call for a real pause option, clear safety triggers, and outside audits speaks to that shared concern across left and right.

What To Watch Next

Watch for four signs. First, independent audits of model autonomy and code-writing impact, not just lab press posts. Second, concrete “stop” rules that pause training when risky skills appear, plus proof those rules are used. Third, verified reports on cyber misuse that show whether models act with less human hand-holding. Fourth, movement in Congress or federal agencies to require test results and safety plans before the biggest models launch.