Safety theater has a new face, and it looks like a benchmark score. AI companies are spending more energy publishing safety metrics than closing the gap those metrics are supposed to measure. The result is a system where models perform well on tests designed to catch the problems they were already trained to avoid â" and deployment reality stays conveniently unexamined.

Third-party verification is the obvious answer, and it's the one nobody does. Independent auditors can't access training pipelines. External researchers work with frozen model weights months after deployment. By the time a safety concern is confirmed, the model is already in production, making decisions at scale. The companies know this. The benchmarks are designed around it.

The Gap Nobody Measures

The distance between a safety claim and a safety outcome is where the real risk lives. A model can score well on refusal-rate tests while becoming increasingly likely to comply with jailbreaks in the wild. Correlation between benchmark performance and field behavior is weak at best, and companies have strong incentives to report the number that looks best rather than the number that tells the truth.

This isn't a technical failure. It's a structural one. Until external auditors have access to active deployments â" not just model weights â" safety scores will reflect what models can be made to do in controlled conditions, not what they actually do when the stakes are real. The PR apparatus of AI safety is now larger than the safety work itself, and that's the consequence of building accountability around self-reporting.