Anthropic CEO Backs Independent AI Watchdogs Amid Safety Concerns

Dario Amodei has become one of the most recognizable voices in the push for tighter AI guardrails, and his case starts with a simple belief: the people building frontier systems should not be left to police themselves. That idea has made him a headline figure in the debate over independent oversight, especially as AI companies race ahead and the risks keep getting more serious by the day.

Amodei, the co-founder and CEO of Anthropic, wants outside watchdogs to play a real role in checking how powerful models are built and tested. One group he points to, METR, presents itself as an AI safety testing lab, but its roots run deep in the same safety-first network that shaped Anthropic’s early culture and many of the people around it.

That network has long overlapped with effective altruism, the movement built around using evidence and reason to do the most good possible. In AI circles, it has become a powerful force, with supporters framing the technology as something that needs careful steering before it starts behaving in ways nobody can control.

METR’s own public pitch is straightforward enough. The lab says it evaluates frontier models so companies and the broader public can better understand what these systems can do, and what kinds of dangers may come with them.

Behind that mission are familiar names. Beth Barnes, METR’s founder and CEO, worked at OpenAI with Amodei during the early days of ChatGPT, and Paul Christiano, who also helped shape OpenAI’s safety work, later founded the first version of METR.

Both Barnes and Christiano have spoken openly in the language of effective altruism, treating AI safety as a moral and practical mission. Barnes once described the goal in stark terms, saying it should be someone’s job to examine models, think through how they might kill people, and spot the warning signs before disaster strikes.

Christiano made a similar case years earlier, arguing that the movement mattered more if it built something genuinely useful than if it simply grew fast. That philosophy helped create a pipeline of researchers, evaluators, and funders who would eventually orbit Anthropic and its flagship model, Claude.

The money trail matters too. Anthropic drew support from some of the biggest names in effective altruism, including Sam Bankman-Fried before his fall from grace and criminal conviction. Jaan Tallinn, the Skype co-founder and a major backer of AI safety causes, also played a leading role in Anthropic’s early financing.

Amodei’s own path helps explain why he sees the issue so sharply. He studied biophysics at Princeton, went on to Stanford Medicine, and then moved through Baidu, Google Brain, and OpenAI, where he climbed to vice president of research.

At OpenAI, he worked during the era when the company was pushing from early language experiments into the models that would later make ChatGPT possible. As the systems became more powerful, Amodei grew uneasy about whether the company was moving fast enough on safety or serious enough about the risks tied to bigger and bigger models.

He eventually left in 2020, saying the environment no longer felt trustworthy. In a recent interview, he said distrust grows when a company’s values, honesty, and motives no longer line up, and that kind of concern made it hard to keep building inside the same walls.

Anthropic was the answer. The company was built around the idea that AI development should be paired with strict limits, and it even gave Claude a kind of constitution, a set of boundaries meant to keep the system helpful without letting it become unsafe, deceptive, or unethical.

That approach has not stayed abstract. Anthropic has clashed with the Pentagon over autonomous battlefield targeting, rejected work it viewed as mass surveillance, and taken a harder line than some rivals when the company believes a project crosses the line.

Amodei wants the rest of the industry to move in that direction, too. His proposal calls for embedded evaluators, third-party experts with ongoing access inside frontier AI companies so they can check safety practices, flag incidents, and look beyond polished public promises.

He argues that the public deserves more than reassurances and marketing language. In his view, outsiders need a real window into what companies are doing, because the stakes around advanced AI are too high to let the builders grade their own homework.

The urgency behind that stance is easy to see. AI systems are getting more capable, more autonomous, and harder to predict, while companies keep learning that even their own agents can behave in ways they did not expect.

Amodei still says the technology can improve human life in huge ways. The catch is that those benefits depend on discipline, patience, and a lot more honesty than the industry has shown so far.