California's SB 813 would establish a framework by January 1, 2028 for selecting and regulating independent verification organisations that assess frontier AI models. METR's use of about 400,000 dollars in API credits for an OpenAI-related review made both the cost of independent scrutiny and the risk of dependence on the company funding access more visible.
California's SB 813 would create a state process for independent organisations that assess frontier AI models. The bill calls for a framework to select and regulate verifiers by January 1, 2028. The measure is intended to specify who conducts outside scrutiny and under which rules, rather than leaving a model developer to rely only on its own safety assessment.[1], [2]
The economic constraint appeared in METR's review linked to OpenAI's Hugging Face incident. The work covered roughly 1,200 agents and 70,000 messages and ran for six days rather than the two planned. METR used about 400,000 dollars in API credits supplied by OpenAI for model access. The amount was computing access from the company under review, rather than a cash payment, sharpening the question of how independent organisations can afford access to expensive models.[1], [2]
In the METR case, OpenAI was the subject of the review, the provider of the model used in the work and the supplier of computing access. That arrangement does not by itself make the findings wrong, but it creates a conflict in which an auditor may depend on the company it examines to continue its work. SB 813's certification framework will therefore have to address technical competence alongside who pays for access and what safeguards protect a verifier's judgement.[1], [2]
