AI Safety or Safety Theatre?

Nearly every major artificial intelligence company now describes safety as a core value.

The word appears in policies, model announcements, regulatory testimony, and public statements. Yet saying that a system is safe is not the same as demonstrating safety.

This distinction matters because AI safety can easily become theatre.

A company may publish principles while withholding meaningful evaluation results. It may add conversational refusals while giving an agent excessive permissions. It may announce a red-team exercise without explaining the scenarios tested, the vulnerabilities discovered, or whether the findings changed the system.

Real safety is less glamorous.

It involves threat modeling, adversarial testing, independent evaluation, incident reporting, data governance, access control, monitoring, change management, and clear accountability. NIST describes AI red teaming as structured testing designed to identify inaccurate, harmful, discriminatory, or insecure behavior. NIST also warns that static benchmarks alone are insufficient because capable adversaries continually discover new attack methods.

Safety must also be evaluated in context.

A model may be safe for drafting routine marketing content but inappropriate for diagnosing a patient. An agent may be useful for summarizing compliance evidence but unsafe if it can approve its own remediation actions.

The same model can therefore present different levels of risk depending on the data, permissions, users, systems, and decisions surrounding it.

I have learned through cybersecurity work that a policy is only valuable when it changes operational behavior. The same principle applies to AI. Safety documents should result in testable controls, measurable outcomes, and responsible decisions.

Leaders should be cautious when safety claims rely primarily on branding or voluntary promises. Independent testing, transparent limitations, documented incidents, and external accountability provide stronger assurance.

The objective is not to embarrass AI companies when weaknesses are discovered. The objective is to create systems that improve because weaknesses are discovered.