AI Safety Evaluator METR Raises $71 Million in Six Months

Five Directions for the Funds

METR explained that the funds will go toward projects across five areas: autonomous capabilities research, tracking recursive self-improvement, evaluating monitoring systems, risk assessment, and AI incident investigation. Recursive self-improvement refers to a scenario in which an AI improves its own algorithms or training process to produce better versions of itself — a concern long considered central to AI safety discussions, given the possibility that this could unfold at an uncontrollable pace. METR said it aims to track exactly how far along this scenario has actually progressed.

Donor List and Independence Principle

The attached image also disclosed a list of donors. The Audacious Project was named as the channel through which METR received its first large-scale institutional funding, alongside the Sijbrandij Foundation, The Pew Charitable Trusts, Schmidt Sciences, the David & Lucile Packard Foundation, the LaCentra-Sumerlin Foundation Frontier Fund, the Astralis Foundation, the AI Security Institute (AISI), Longview Philanthropy, and the Survival & Flourishing Fund. Individual donors mentioned included people affiliated with Jane Street, along with David Farhi, Geoff Ralston, Dylan Field, and Steve Newman.

In the tweet, METR stated, "We work to maintain independence from frontier AI companies," adding that it "does not accept funding from these companies, nor donations made in the name of or at the direction of their employees." This appears to reflect a judgment that, as METR's risk assessments carry increasingly greater impact, accepting funding from the very companies it evaluates could raise conflict-of-interest concerns.

Why This Principle Matters

Within the AI safety evaluation ecosystem, a pattern has emerged in which frontier model developers like OpenAI and Anthropic maintain their own internal evaluation teams while also granting outside organizations like METR pre-release access for risk assessments. For this structure to earn trust, evaluating bodies need to be financially independent from the companies being evaluated. The fact that METR chose to explicitly reaffirm its independence policy in this announcement suggests a recognition that preserving this trust becomes more important as the organization grows in scale.

Meanwhile, METR said it will use the funds to expand its team and launch new projects, and posted a link to job openings alongside the announcement. The exact scale of hiring or a detailed list of new projects was not disclosed.

Editor's Take

What stands out in this announcement is less the dollar figure itself than the funding structure. METR is not an AI company but an organization positioned to evaluate them, and most of the $71 million it raised came from foundations and individual donors. Compared to the tens of billions of dollars that companies like OpenAI or Anthropic raise in funding rounds, the sum is modest — but the nature of this money is entirely different. An evaluating body publicly committing to refuse funding from the entities it evaluates is analogous to an auditing firm striving not to depend on consulting revenue from the companies it audits.

Anyone who has followed the AI safety evaluation industry will recognize how much the landscape has shifted in just the past few years. As recently as two or three years ago, risk assessment of AI models was mostly handled by developers' internal teams, or, even when outsourced, often tied to the developer through contractual relationships. The fact that an organization like METR can explicitly declare independence from any single company by pooling funding from a diverse group of independent foundations signals that this field is establishing itself as a distinct industry in its own right.

Organizations in Korea working on AI governance or policy may find this model instructive. A structure in which an institution seeking to conduct AI safety evaluations or audits receives direct commissions from developers may be stable in the short term, but over the long run it risks undermining the credibility of its own assessments. METR's case demonstrates that diversifying funding across multiple foundations and individuals, while more cumbersome to set up initially, is ultimately the path that protects the credibility of evaluation outcomes.

In the coming months, the key thing to watch will be what kind of assessment results METR produces with this funding. Given that the organization has explicitly framed tracking recursive self-improvement as one of its priorities, METR's name is likely to appear more frequently in risk assessment reports each time a next-generation frontier model is released.