AI answers may need a clearer ‘I don’t know’ signal
AI tools often give confident answers even when they are unsure. Adding an “I don’t know” button may not fully solve this, because current models may not clearly know what they do and do not know. A practical workaround is to add that tell the AI to admit when its is low.
A stronger check is to ask the model for many separate answers and compare them. If the answers vary a lot, the model may be guessing across a gap in its knowledge. If the answers are mostly the same, the answer may be more reliable.
This kind of checking can use 10 to 100 times more tokens, so it can become expensive. Some people also want AI to keep trying on hard problems instead of giving up too early, so a low- warning outside the answer may be better than a blunt refusal.
Key points
- The core problem is AI giving confident answers when it should admit .
- An “I don’t know” button may not work if the model cannot reliably judge its own limits.
- can push the AI to say when its is low.
- Comparing many answers can reveal , but it may use far more tokens.
- For hard creative work, a low- warning may be better than making the AI stop too soon.