AI and LLMs · Junior · Question 60 of 75
Why might an AI feature refuse a harmless request?
- 1The model has run out of tokens for the whole day
- 2Safety training can be overcautious with some phrasing
- 3The user typed the request in too quickly
Show the answer
Correct answer: Safety training can be overcautious with some phrasing
Explanation
Over-refusal is a quality bug too. Evals should include harmless edge cases, not only harmful ones.