On Sept. 9, METR announced a formal agreement with Anthropic to conduct an “independent investigation of agent incidents” involving the company’s large‑language models, and three days later Redwood Research disclosed that several of its staff had been subcontracted to METR for the same probe.
METR, formerly known as ARC Evals, spun off as a separate nonprofit in 2023 and is backed by prominent figures in the Effective Altruism (EA) community. The EA‑aligned fund Coefficient Giving – co‑founded by Holden Karnofsky and Facebook co‑founder Dustin Moskovitz – has poured more than $100 million into METR‑related work, including a $1.5 million grant to METR’s incubator, the Alignment Research Center, in 2022.
Redwood Research, another leading AI‑safety nonprofit, also enjoys deep ties to Anthropic. Redwood received a $36 million grant from Coefficient Giving in November 2023, followed by a recommendation for an additional $70 million to “scale up” its work after a high‑profile Hugging Face investigation. Redwood’s original board included Karnofsky, who is an Anthropic employee and married to CEO Dario Amodei’s sister, Daniela Amodei, as well as Paul Christiano, a former OpenAI colleague of Amodei and a trustee of Anthropic’s Long‑Term Benefit Trust.
Industry observers argue that the web of personal, financial and governance connections makes it impossible for these groups to act as truly independent overseers. Perry Metzger, chairman of the Washington‑based Alliance for the Future, said, “None of these people are independent, none of these people are arm’s length.” Congressional critics echoed the sentiment, with Rep. Josh Gottheimer (D‑NJ) warning that AI firms’ eagerness to adopt self‑policing frameworks should raise red flags, and Rep. Steve Scalise (R‑LA) dismissing the arrangement as “the people we’re trusting to beat China in AI? Give me a break.”
Representatives for METR maintain that the nonprofit does not accept compensation from AI labs and that its work is guided by conflict‑of‑interest policies. Redwood’s CEO Buck Shlegeris said the organization’s collaborations with AI companies have largely been research‑focused rather than external accountability work, and that Redwood follows METR’s conflict‑avoidance rules when the two partner. Anthropic and OpenAI did not immediately respond to requests for comment.
The controversy highlights a broader debate over how the fast‑growing AI sector will be regulated. Lawmakers on Capitol Hill are watching closely as companies like Anthropic and OpenAI seek to shape oversight mechanisms that could preempt stricter federal legislation, while Pentagon officials have publicly criticized what they describe as “death‑cult‑like philosophies” driving such self‑regulatory efforts.