Anthropic and OpenAI have proposed allowing independent third-party evaluators to monitor their AI development processes from within the companies.

Anthropic CEO Dario Amodei has proposed a plan to embed third-party safety evaluators within frontier AI companies to monitor the development of large language models. OpenAI CEO Sam Altman has also committed to this practice. The proposal aims to provide independent oversight by granting evaluators access to internal systems and the ability to publish findings, following concerns about the rapid pace of AI development.

Proposed Oversight Model

Under the proposal, independent groups such as METR and Redwood Research would receive access comparable to internal risk teams. Amodei compared the model to banking industry supervisors who maintain a continuous presence within financial institutions. Evaluators would be permitted to report findings publicly, subject to limited redactions, without requiring editorial approval from the AI companies.

Limitations and Industry Skepticism

Legal experts and researchers have raised concerns regarding the effectiveness of the plan. Julie Andersen Hill, an expert on banking regulation, noted that unlike bank examiners, these AI evaluators lack the legal authority to force management changes or halt model training and releases. Other researchers emphasized that for the system to be effective, it must be backed by legislation to ensure evaluators function as independent watchdogs rather than company-controlled vendors.

Access to Training Data

Evaluators argue that access to final models is insufficient, as models may conceal problematic behavior during testing. They propose monitoring intermediate training checkpoints to identify when concerning behaviors emerge. This would allow for verification of company claims regarding training processes, such as whether a model attempted to undermine its own alignment training.