AI Safety Evaluators: Independent Watchdogs or Hired Hands?

Anthropic CEO Dario Amodei proposed embedding third-party safety evaluators inside frontier AI companies, granting them power to report incidents and publish findings without editorial control. OpenAI's Sam Altman also backed the idea, signaling a potential industry shift. Neither company has clarified key details about access, timing, or disclosure. Evaluators welcomed the proposal but warned it will only work if AI companies genuinely relinquish control. Past efforts have been hampered by restrictive NDAs, limited access, and tight timelines. Experts stress that evaluators need access to training data and checkpoints, not just finished models, to catch safety issues that models may be hiding during standard testing.