Anthropic and OpenAI propose embedded independent AI evaluators
Anthropic CEO Dario Amodei has proposed placing third-party safety evaluators inside frontier AI companies, with access to systems, authority to report incidents and the ability to publish findings without company editorial control. OpenAI CEO Sam Altman said his company would also commit to the practice. Researchers broadly welcomed the idea but said its details remain unresolved.
Evaluators want access beyond released models, including training checkpoints, post-training environments, transcripts and logs. They say this could reveal when concerning behavior emerged and help detect models that recognize tests and behave safely only under evaluation. Apollo Research’s Alexander Meinke said companies should be able to show that models did not try to undermine alignment training, while Far.AI’s Adam Gleave said interviews with employees could help verify public safety claims.
Independence may be difficult under restrictive confidentiality agreements, limited time and company control over publication. METR and Redwood Research had about a week to examine an OpenAI incident, while Apollo had three days to test GPT-6 Astra and said the results were inconclusive. Researchers called for public standards and regulation. California laws and the EU AI Act already include some independent assessment requirements, but they are narrower than the proposed arrangement. Meta, SpaceXAI and Google DeepMind have not committed to embedded evaluators.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.