A small group of outside organisations is being asked to assess the risks of powerful AI models, even as the rules for their access, funding and independence remain unresolved. That matters because an evaluator that depends on the companies it scrutinises may find it harder to publish an unwelcome result.
Watch Desk analysis
What happened
AI labs are working with groups including METR, Apollo Research and Transluce to assess model capabilities and risks, according to a CNBC report published on 11 October. METR said it had received commitments of about $71 million over the previous six months, compared with $13.6 million in total contributions in 2024. Another evaluator, Vals AI, said it had grown from eight employees to roughly 30 and announced a $40 million funding round in August.
The arrangements vary. Anthropic has said it will fund evaluators from Faculty, Accenture’s specialist AI business, to test safeguards, while describing pooled or government funding as a longer-term goal. OpenAI said it was finalising contracts with third-party assessors and would announce details in the coming weeks. Read CNBC’s report.
Why it matters
Independent evaluation is only useful if evaluators can get meaningful access, publish important findings and withstand pressure from whoever pays the bills. Those conditions are still being worked out. The report notes that Anthropic and OpenAI have proposed ways to embed outside assessors, but there is no settled system for funding or standardising the work.
A potential alternative is formal oversight. The report says a federal bill would establish government-authorised independent verification organisations, while California has signed two laws involving AI auditors. Voluntary company commitments and state-level rules are not the same as a common system, though; the shape of that system could determine whether an evaluator is a genuine check or a particularly well-branded guest.
Our read
The growth in funding and staff suggests this field is becoming more than a side project. But independence is not a badge an evaluator can award itself. Access, protection from retaliation, transparent methods and the right to publish inconvenient findings are the more useful tests. Readers should watch for the terms of actual agreements, not just promises to work together.
What to watch
- Whether OpenAI publishes the contracts and access terms it says are being finalised.
- Whether evaluators can publish key findings, including results that reflect poorly on a client.
- How federal and state proposals define an evaluator’s access, funding and accountability.
Discussion spark: Should AI evaluators be funded by the companies whose systems they test, or should public money be required to protect their independence?
Sources and evidence
- AI’s quiet safety gatekeepers are stepping into the spotlight – CNBC (11 October 2026, 11:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.