Who is Dario Amodei, the Anthropic CEO pushing independent AI oversight?

Anthropic CEO Proposes Independent Evaluators for AI Development Oversight

Anthropic chief executive Dario Amodei has called for frontier AI companies to grant ongoing access to independent third-party evaluators to verify compliance with safety practices during model development. In a recent public letter, Amodei described a system of “embedded evaluators” that would receive employee-like access to training pipelines, completed models, and internal processes at participating firms. He cited METR as one example of an organization capable of performing such evaluations.

METR operates as an AI safety testing laboratory that assesses frontier models for capabilities and potential risks. Its work includes evaluations of systems developed by multiple companies, including Anthropic’s Claude model. The organization’s website states that its goal is to help companies and society understand AI capabilities and associated risks. Amodei’s proposal would require companies to commit to incident reporting and alignment assessments conducted by such external teams, with the stated aim of increasing public information about development decisions.

Amodei’s background includes a Ph.D. in biophysics from Princeton University in 2011 and postdoctoral work at Stanford University School of Medicine. He later held research positions at Baidu, where he contributed to machine-learning speech recognition systems, and at Google Brain, where he worked on neural network research. In 2016 he joined OpenAI, rising to vice president of research and participating in the development of early GPT models. He left the company in 2020.

Amodei has stated that his departure stemmed from concerns over whether OpenAI would prioritize safety measures and maintain alignment between stated values and operational decisions. In a 2026 Bloomberg interview, he described difficulties trusting an organization when patterns of behavior appeared inconsistent with public commitments. After leaving, he co-founded Anthropic in 2021 with a stated focus on AI safety.

Anthropic has published model cards, safety reports, and a written constitution for its Claude model that specifies boundaries for behavior. The constitution outlines requirements that the model remain helpful while avoiding unsafe, unethical, or deceptive actions. The company has also released documentation detailing alignment techniques applied during training. These materials constitute public records of the firm’s internal safety processes.

Anthropic has declined certain contracts on safety grounds. It turned down a proposed $200 million Department of Defense agreement involving tools for autonomous targeting and declined work it assessed as enabling mass surveillance. Amodei has said these decisions reflect the company’s view that some applications require stricter controls, particularly as other developers have reported difficulties managing autonomous agents.

Amodei’s letter emphasizes that AI capabilities can produce broad benefits if developed with deliberate safety standards. He wrote that the technology can improve quality of life provided companies build it “in the right way” and take “unusually deliberate care.” The proposal for embedded evaluators is presented as one mechanism to make development decisions more transparent to the public.

METR has conducted evaluations involving Anthropic models. Its founder and CEO, Beth Barnes, previously worked at OpenAI during the period when early GPT systems were under development. Paul Christiano, who established an earlier version of the organization, also contributed to research on model behavior at OpenAI. METR’s evaluations have examined questions of model risk and capability.

Anthropic’s early funding included participation from investors such as Sam Bankman-Fried through FTX in 2022 and Jaan Tallinn in 2021. Bankman-Fried was later convicted in a fraud case involving FTX. Tallinn has supported organizations focused on existential risk research. Anthropic has also received investment from other sources as the company scaled.

Amodei has compared aspects of AI development competition to historical arms-control discussions, suggesting the possibility of negotiated limits between major actors. He has continued to advocate for industry-wide safety commitments while maintaining that AI progress can proceed under appropriate safeguards. Anthropic has allocated funds, including a reported $20 million donation to a political advocacy group supporting candidates favoring AI safety regulations.

The company’s public materials describe Claude as designed to assist users and society while adhering to defined constraints. Amodei has reiterated that achieving safety benefits requires time and structured evaluation processes rather than rapid unchecked scaling.