Independent AI Evaluators Face Funding and Independence Challenges as Major Labs Adopt Embedded Models

As the AI industry grows into a multi-trillion-dollar sector, independent evaluators like METR, Apollo Research, and Vals AI have become critical watchdogs. With federal regulation absent, major companies including Anthropic, OpenAI, Google, Meta, SpaceX, and Nvidia signed a "morally binding" agreement in September supporting self-policing and cooperation with external auditors.

Anthropic and OpenAI are leading the shift toward "embedded" evaluation. Anthropic plans to place staff from Accenture's AI division inside the company to test safety guardrails, with Anthropic directly funding the effort. OpenAI is finalizing contracts with independent safety organizations like METR and Redwood Research, though it recently fired three employees allegedly for communicating with third-party evaluators, a move the company denies was related to safety whistleblowing.

Funding and Structural Conflicts

The financial sustainability and independence of these evaluators remain contentious. METR, a non-profit, secured approximately $71 million in funding commitments over the last six months, a sharp increase from $13.6 million in total donations in 2024. Vals AI, a for-profit benchmarking startup, raised $40 million in August and expanded its team from 8 to about 30 employees.

Experts warn that direct funding from the companies being evaluated creates structural conflicts of interest. Suresh Venkatasubramanian, a Brown University professor, noted that auditors may face business pressure if they issue unfavorable reports. Kevin Werbach of the Wharton Accountable AI Lab pointed out that the ecosystem is not yet robust, with METR having fewer than 50 full-time employees.

Regulatory and Legislative Moves

Legislative efforts are emerging to formalize independent verification. The FRONTIER Act, introduced in July by Rep. Lori Trahan and Rep. Jay Obernolte, includes provisions for Independent Verification Organizations (IVOs). Fathom launched a marketplace framework for IVOs in June of last year, and California Governor Gavin Newsom signed the first state-level IVO framework and auditor registry in September.

The AI Evaluator Forum, including METR and AVERI, published a letter in September outlining "Minimum Conditions for Embedding Evaluators." The letter demands that embedded evaluators receive access privileges equal to highly privileged employees and protection from retaliation. It explicitly states that embedded evaluation is a complement to, not a replacement for, external oversight.