Anthropic's AI doom warnings hit Washington as Trump and Johnson resist a slowdown

How the latest round of AI alarm started

The current debate began with a resignation. According to TechCrunch, researcher Jacob Coxon, who previously worked at OpenAI, announced he was leaving Anthropic because he believes the leading AI labs are "gambling with our lives." The post gained immediate traction when Anthropic's alignment lead shared it on X, writing "We really do earnestly believe AI could kill all humans!" and adding a personal estimate that the chance of that outcome is greater than 10 percent within the next decade.

The timing made the moment combustible, as TechCrunch's podcast panel noted: it followed recent model releases from Anthropic and OpenAI's Astra, a security incident involving an OpenAI internal model on Hugging Face, and growing reporting about OpenAI internal agents accessing wikis and leaving messages for each other.

The TechCrunch Equity hosts were largely skeptical of the doom framing. Anthony Ha questioned who the "we" speaking for the industry really is and dismissed the 10 percent figure as a number pulled from the air, while crediting Coxon as someone who actually put his career behind his convictions rather than voicing alarm while continuing to build. Kirsten Korosec asked whether catastrophe warnings double as a strange kind of capabilities flex — proof the models are advanced enough to be dangerous — especially with major labs preparing to go public. Sean O'Kane raised a more concrete question: with Anthropic's S-1 filing reportedly weeks away, its lawyers may now have to translate employees' public extinction estimates into formal risk-factor language.

Amodei calls for pacing the frontier

Days into the controversy, Anthropic CEO Dario Amodei published a lengthy open letter arguing it is time to "pace the frontier" and deliberately slow AI development, The Verge reports. The proposal produced an unusual show of agreement among rivals: OpenAI's Sam Altman and Elon Musk publicly voiced support on X, and Alphabet's Demis Hassabis offered tentative backing.

Washington pushes back

That consensus stopped at the political level. According to the Financial Times, cited by The Verge, President Trump dismissed the executives' concerns, saying "we're leading China in AI ... and, frankly, I want to keep it that way, because whoever wins AI, wins."

House Speaker Mike Johnson echoed the view on CNN, telling Jake Tapper that rushing to restrict AI development could itself be a "national security threat" and urging people not to panic. "If Congress just races in and does some sort of emergency session to try and regulate AI, we will lose the race to China," he said.

The Republican position, as The Verge frames it, rests on a single fear: that any pause or regulatory friction gives China the edge in the AI race.

The evidence on fast timelines is contested

The political fight is happening while the technical case for imminent danger remains unsettled. MIT Technology Review reported on a multi-institution study led by Peter Kirgis and Sayash Kapoor at Princeton that tested whether AI agents can do open-ended AI research, the skill underlying recursive self-improvement.

Using a method the researchers call "shadow evaluation," they had Anthropic's Claude Opus 4.8, running on open-source software called OpenClaw, tackle research questions drawn from two unpublished NeurIPS 2026 papers. Given six days, $3,000 in Anthropic API credits, a GPU budget and web access, the agents completed all the engineering — literature reviews, hundreds of experiments, a written paper — but the original authors rejected both outputs. The agents, Kapoor said, were "unambiguously bad" at the research itself: they locked onto unpromising approaches, could not rethink them from scratch, and made no novel contribution. A follow-up run with Anthropic's newer model Mythos, which the Trump administration has restricted to approved organizations, is underway.

Anthropic cofounder Jack Clark wrote that the lack of creativity in current systems is a "bearish signal on short recursive self-improvement timelines" — a view that sits awkwardly alongside both apocalyptic forecasts and Washington's race rhetoric.

Why it matters

The episode marks the first time the industry's existential-risk argument has been met with a direct, coordinated political counterargument, and the China-race framing now shapes what regulation is politically possible in the US.

It also creates disclosure and credibility problems for the labs themselves. If Anthropic proceeds toward an IPO while its leadership calls for slowing down and its staff publicly estimate double-digit odds of catastrophe, the gap between stated beliefs and business conduct will be scrutinized by investors and skeptics alike.

Finally, the Princeton results cut against both sides' certainty: if automating AI research is further off than the boldest forecasts suggest, the case for immediate emergency slowdowns — and for panic about losing a race — both rest on shakier ground than either the doomers or their Republican critics admit.