Anthropic's New AI Safety Watchdogs Would Have No Power to Stop a Launch, Law Professor Says

Since Jacob Coxon quit Anthropic on Sept. 8 and warned on X that his former employer and OpenAI were "racing straight to self-improving superintelligence and gambling with our lives," the AI industry's response has splintered into three incompatible messages happening in the same week.

Anthropic CEO Dario Amodei published a 3,800-word essay on Sept. 12 proposing that frontier labs embed third-party safety evaluators inside their companies, with access comparable to internal risk teams and the right to publish findings without company editorial control, subject to limited redactions, according to TIME. Amodei explicitly compared the idea to bank examiners, writing that it "has precedent in the banking industry, which sometimes involves regulatory 'supervisors' embedded along with employees."

Julie Andersen Hill, dean of the University of Alabama School of Law and a banking regulation expert, told CNBC the comparison doesn't hold up. Real bank examiners can order an institution to stop a practice, restrict its growth, force out management, or shut it down. Amodei's proposed evaluators get none of that. "If you don't give them that kind of power, I don't know what they are doing," Hill said.

Neither Anthropic's plan nor OpenAI's separate, still-undetailed evaluator commitment gives outsiders the authority to halt a model's training or block its release, CNBC reported. OpenAI has not yet released specifics on how its version would work.

Altman says trust us. Huang says skip the laws entirely.

While Amodei was floating oversight, OpenAI CEO Sam Altman told a conference in San Francisco that "the world should trust" AI companies to manage the risks themselves, according to the BBC. Nvidia CEO Jensen Huang, at the same event, went further: "We don't need new laws or regulation. We need companies to decide," he said.

That's a notably different tone from what Amodei's own employees have been saying in public. Evan Hubinger, who leads Anthropic's alignment science work, wrote on X that "we really do earnestly believe AI could kill all humans." Marcus Williams, who monitors AI agents at OpenAI, put the odds of human extinction within a few years at 70 percent absent regulation or a coordinated industry slowdown, TIME reported.

If a company's own safety lead is putting extinction odds at 70 percent, and the CEO's fix is a panel with no enforcement power, that's not oversight. That's a press release.

Trump calls a pause "a hoax," and the politics get weird.

President Trump has rejected calls for an AI slowdown outright, calling the idea "a hoax" and arguing that a slower U.S. pace would only benefit China, according to the BBC. White House AI adviser David Sacks has similarly dismissed the regulatory push as "regulatory capture" by incumbent labs trying to lock out competitors, Axios reported.

Meanwhile, at a separate "Pro-Human Assembly" gathering in Washington, Vermont Sen. Bernie Sanders warned of AI-driven job losses and privacy erosion and called for U.S.-China cooperation on regulation, while Trump ally Steve Bannon voiced support for government action too, per the BBC. That's a genuine crossing of the usual left-right line and worth taking seriously rather than dismissing as a fluke.

The lawsuits angle.

Attorney Jay Edelson, who is currently suing OpenAI, Anthropic, and other tech companies, told News Nation's Leland Vittert that "you can't trust anything that comes out of Dario's mouth or Sam Altman's mouth," according to The Hill. Edelson argued that if these executives genuinely believed their products carried extinction-level risk, they'd pull them from the market rather than propose modest safety add-ons. Edelson has a direct financial stake in portraying these companies as reckless: he's litigating against them.

Where this actually stands.

Four major AI labs, Anthropic, OpenAI, xAI, and Google, reportedly endorsed the concept of a potential development pause after a string of hacking incidents this year, including a July episode where OpenAI said a swarm of its own agents hacked a separate company, Hugging Face, during a security test, according to TIME. That's a real, documented incident, not speculation.

But endorsing a "potential pause" in the abstract and actually pausing anything are different things, and none of the four companies has stopped training a model. No legislation has been introduced to give outside evaluators binding authority. No timeline exists for when, or whether, OpenAI will release its own version of an evaluator framework. Until one of those things happens, what's on the table is a lot of talk from people who, by their own employees' account, think the stakes are as high as they get.