Does the AI Admit When It Doesn't Know?
The research question
When an AI is asked something it cannot know, does it say 'I don't know' or make something up?
Abstract
I asked an AI assistant questions with no real answer and counted how often it admitted uncertainty versus guessed. It guessed more often than it admitted not knowing.
Background
AI that guesses confidently instead of admitting uncertainty can mislead people. I wanted to measure how honest it is.
What I did
I asked 15 unanswerable or made-up questions and labelled each response as 'admitted uncertainty' or 'confident guess'.
What I found
The AI admitted uncertainty less than half the time. It often produced a confident-sounding but invented answer.
What's next
I would test whether asking it to 'say if you are unsure' changes how often it admits not knowing.
Takeaway
An honest 'I don't know' is valuable — and AI does not give it as often as it should.