OpenAI and Anthropic are pushing to embed independent safety assessments within their labs: assessors will have greater access to the training process and model behavior, and will be able to report anomalies when issues are identified. This shift from external rhetoric to institutional design remains unclear regarding the scope of access, the timing of intervention, the right to publish results, and whether assessors can truly operate independently of company control. Supporters argue this moves oversight to the training phase; skeptics worry that without adequate rules and legal protection, it will ultimately become a controlled facade.
