AI
Hallucination
Also called
- Confabulation
A language model predicts plausible continuations. Plausibility and truth usually coincide, which is why these systems are useful — but when they diverge, the model has no internal signal telling it so. The wrong answer arrives with exactly the same confidence as the right one.
This is a property of the technology, not a bug awaiting a patch. The engineering response is not to eliminate it but to build around it: ground the model in retrieved sources, ask it to cite them, score its own confidence, and route anything below a threshold to a human.
The risk is proportional to how expensive the action is to reverse. A wrong summary someone reads is cheap. A wrong value written into a claims system is not. That asymmetry is what should decide where the human sits in the loop.
Hallucination risk is the reason an AI deployment needs a confidence threshold and a review path — the technical design follows directly from what a wrong answer would cost you.
Commonly misunderstood
What people get wrong
The claim
“The newer models don't hallucinate.”
What is actually true
They hallucinate less and more convincingly. Lower frequency with higher fluency can be worse operationally, because reviewers relax exactly as the errors become harder to spot.
Where this comes up
Services where it matters
Next step
Working through a hallucination decision?
Tell us the situation. We will give you the tradeoffs as we see them, including when the answer is that you do not need what you are being sold.
No pitch deck. A 30-minute conversation about what you are trying to achieve.