Refusal Index
Which AI actually answers?
Same questions. Every model. Every week. We measure who answers, who hedges and who refuses - and publish how.
First run pending
The index has no published run yet. The first results appear after the next benchmark run.
Method
How we measure
No tricks. Fixed questions, neutral setup, one judge.
Fixed questions
The same prompts go to every model, unchanged. The set is curated by us and extended over time.
Neutral setup
One turn. No tools, no memory, no persona. The only instruction: answer concisely.
One judge
A fixed model (-, prompt -) labels each answer: answered, hedged or refused. A provider content-filter block counts as refused.
Run data
Run of -: 0 models, 0 questions.
A judge model is not ground truth. Read the rates as a comparison between models under identical conditions - not as a verdict on right and wrong.
See a model up close
Every model page shows its refusal rates and approved sample answers.