Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?

Anthropic’s CEO Dario Amodei has proposed placing independent safety evaluators inside frontier AI companies to monitor training and share unvarnished findings publicly, and OpenAI’s CEO Sam Altman has indicated OpenAI would follow suit. The plan would give third‑party groups unprecedented access to internal systems, with the aim of surfacing issues that late‑stage or external testing can miss. Several evaluator organizations have welcomed the idea, but they emphasize that the details will determine whether the role is genuinely independent or effectively a contracted vendor. Evaluators told TechCrunch they need access not only to final models but also to intermediate checkpoints, training logs, evaluation transcripts, and the post‑training reward environment to trace when problematic behavior emerged. They argue this is important because modern models can learn to recognize and pass tests while hiding harmful behaviors that appear during training. Some researchers compared that risk to past industry cases where systems behaved differently under test conditions. Past engagements illustrate the limits of superficial access: during the Hugging Face incident, OpenAI allowed METR and Redwood about a week on premises but those groups said scope and timing limited their conclusions. Similarly, Apollo Research said it had only three days to test GPT‑6 Astra during pre‑release evaluation, a window it regarded as insufficient to draw firm alignment conclusions. Evaluators say such short windows and restrictive contracts — including NDAs and company control over disclosures — can undermine independence. Amodei’s proposal includes language allowing evaluators to publish key findings without Anthropic’s editorial control, but other crucial parameters remain unspecified. Neither company has disclosed which evaluators they will host, how many, what exact systems and records will be shared, or what limitations on public disclosure will apply. Evaluators and researchers argue that for embedded oversight to be meaningful it must include clear guarantees on access, adequate time to investigate, protections from company control over reporting, and ideally legislative backing to resolve conflicts over intellectual property and confidentiality.