AI refusals are unreliable, opaque and costly, the article argues

Key points

• Companies train models to refuse prompts, using other AI models to run exercises that reward refusals and penalize over-refusal.
• Anthropic reportedly said one type of classifier added 24% to its chatbots' compute costs earlier this year.
• Refusals are probabilistic: repeated identical risky questions were generally refused, but every so often a model answered, per Ryan McBain's tests.
• AI companies currently decide where the refusal line falls, and they reportedly do so with utmost secrecy.

Why it matters for marketers: Content teams should expect inconsistent refusals on borderline topics and should not assume a tool's safety filters are fixed or fully predictable.