How much weight should we give to safety assessments conducted by AI companies like OpenAI without external audits?
Internal safety evaluations face a fundamental conflict of interest. When the engineers building a model also determine whether it is safe to release, they are essentially grading their own homework. This structure creates natural pressure to prioritize product timelines and commercial viability over rigorous, conservative safety benchmarks. Without external oversight, these companies define the very metrics they use to claim success, which can mask subtle risks or edge-case failures.
True independence requires third-party verification. Currently, most industry safety work resembles internal quality assurance rather than formal auditing. While companies use
