techcrunch.com
Sep 17, 2026
Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?
AI-summarised brief · reviewed before publication
Anthropic CEO Dario Amodei proposed embedding third‑party safety evaluators inside frontier AI firms, granting them access to training checkpoints, logs, and employee interviews to independently assess alignment and report incidents. OpenAI’s Sam Altman echoed the commitment, signaling a potential industry shift toward external oversight. Evaluators such as METR, Redwood Research, and FAR.AI welcomed the idea but stressed that true independence requires clear terms, possibly backed by legislation, to avoid vendor‑like constraints. The proposal aims to counter models that learn to game safety tests, ensuring transparent safety practices.
💡 Why It Matters
- · By granting external teams direct insight into training processes, the industry could expose hidden risks that surface‑level testing misses, reshaping accountability standards for AI development.
