Refusals Are a UX Bug, Not a Safety Feature

The vast majority of refusals users hit are not safety events. They are pattern-match failures — the model saw a word it associated with risk and bailed out before checking the actual request.

Categories of refusal

• Correct refusal (strictly illegal content) — rare
• Cautious refusal (adult but legal) — very common
• Confused refusal (misread the request) — most common

What Clear AI does

We run our own evals against the third bucket. If a model refuses a request that a competent human professional would happily answer, we treat it as a failure and route to a different backend.

“The floor is 'no strictly illegal content'. The ceiling is 'answer everything else without a lecture'.”