A hallucination is model output that sounds confident and fluent but is false or unsupported by any source: invented citations, fake API methods, wrong dates.
The model is trained to produce likely text, not verified facts. Knowledge sits compressed in the weights, so gaps get filled with something plausible.
Post-training rewards confident answers, so guessing often scores better than admitting ignorance.
Nothing in the API flags a hallucination. Grounding and verification stay the application's job.
Rewriting in plainer words…
This answer doesn't lend itself to a diagram - it reads best . No credits were charged.
Saved in this browser - sign in to keep your review list.
How should your speech become text?
Listening… your words appear above as you speak - tap Stop when you're done.
Recording · cr - tap Stop & transcribe when you're done.
Transcribing with AI…
Voice:
Keep going - a few more words and AI can grade it.
Interview lens
How interviewers actually use this question
Likely follow-ups
Why doesn't setting temperature to 0 stop hallucinations?Temperature only changes how the next token is picked. If the most likely continuation is wrong, greedy decoding returns it every time.
If you ground the model with RAG, can it still hallucinate? How?Yes. Retrieval can miss or return the wrong chunk, and the model can still ignore or misread it. Name one check, such as verifying citations against sources.
Can token log probabilities tell you when the model is making something up?Low probability can flag uncertainty, but models are often confidently wrong and post-training can skew calibration. It's a signal, not a detector.
A hallucination is output that sounds confident and fluent but is false or has no source behind it, like an invented citation, a fake API method or a wrong date.
It happens because the model is trained to produce likely text, not verified facts.
Its knowledge sits compressed in the weights, so when there's a gap, it fills it with something that looks plausible.
Post-training adds to the problem, because it rewards confident answers, and guessing often scores better than admitting it doesn't know.
Nothing in the API flags a hallucination, so grounding and verification stay the application's job, not the model's.
Weak answers to avoid
Calls it a bug the next model will fixIt follows from training the model to predict likely text. Say you can reduce it with grounding and checks, but you can't expect it to go away.
Treats temperature 0 or RAG as a cureGreedy decoding still returns a wrong answer if it's the most likely one, and retrieval can fail. Present these as ways to reduce it, not guarantees.
Defines it but never explains whyThe question has two halves. Give the causes too: likely text over verified facts, gaps in compressed knowledge, and post-training that rewards confident guesses.
Says the model is lying or knows it's wrongThere is no intent. The model produces plausible text either way and nothing in the API marks the false parts, so the application has to verify.
Interview lens
Likely follow-ups, what you can say, and the weak answers to avoid.
