Dario Amodei calls for embedded AI watchdogs as the frontier safety debate sharpens

The question of who should scrutinize frontier artificial intelligence has become a live industry dispute, with companies reporting difficulties controlling their own agents. Anthropic CEO Dario Amodei has entered that debate with a concrete proposal: give independent third-party evaluators ongoing, employee-like access to AI development pipelines from the inside.

In a recent letter, Amodei proposed that each frontier AI company commit to embedded evaluators whose role would cover safety practice adherence, incident reporting, and alignment assessment across training processes and completed models. He cited METR, an AI safety testing laboratory that describes itself as evaluating frontier models to help companies and society understand AI capabilities and associated risks. "Regardless of what commitments we make, the public deserves to know what is going on. We are still the ones choosing what to include and omit. Embedded evaluators will change this dynamic," he wrote.

The network behind the proposal

The sector-wide significance of Amodei's proposal is complicated by the connections between METR and the community that helped capitalize Anthropic itself. METR's founder and CEO, Beth Barnes, worked at OpenAI alongside Amodei during the development of early ChatGPT versions. Paul Christiano, who led OpenAI's research on ensuring models followed acceptable output strategies, founded the first iteration of METR. Both Barnes and Christiano have framed their work through the lens of effective altruism, the evidence-and-reasoning movement that shaped a significant part of Anthropic's early investor base, and both have worked on evaluations involving Anthropic's Claude models.

That investor base includes figures whose later trajectories complicated the EA story. Sam Bankman-Fried, then one of effective altruism's most prominent public advocates, led Anthropic's 2022 Series B financing round before his cryptocurrency exchange FTX collapsed and he was convicted in a multibillion-dollar fraud case. Jaan Tallinn, the Skype co-founder who has spoken at Effective Altruism Global conferences and donated more than $1 million to the Machine Intelligence Research Institute, led Anthropic's 2021 Series A.

Amodei's path to the safety argument

Amodei came to AI governance through research. He earned a PhD in biophysics from Princeton in 2011, then held a postdoctoral position at the Stanford University School of Medicine before moving through roles at Baidu, where he worked on speech recognition via machine learning, and Google Brain, where he focused on neural network development and safety. He joined OpenAI in 2016, eventually becoming vice president of research, before leaving in 2020 over concerns that the company was not installing adequate guardrails or acting with the honesty he required of it.

Anthropic's own record on safety has carried a concrete cost. The company declined to develop tools the U.S. Department of Defense sought for autonomous battlefield targeting and refused work it considered mass surveillance. That position, Amodei has said, cost Anthropic a $200 million contract. The broader read-through for his embedded evaluator proposal is structural: the organizations he envisions filling that watchdog role are woven into the same professional and ideological network that funded the company they would be tasked with watching.