Related question
Why does the same AI tool give a different answer to the same question?
Short answer
Because it predicts rather than retrieves. A language model constructs each answer word by word based on probability and pattern, so the same input can produce different output. That is a property of the technology, not a defect, and the remedy is to constrain the request.
This surprises executives more than almost anything else, and it undermines trust quickly when nobody explains it.
These systems do not look up an answer and hand it to you. They generate one, predicting what should come next based on patterns in the material they were trained on. That is math, not comprehension. It feels like understanding because the prediction is very good, but it is a different thing.
Once you grasp that, several practical conclusions follow. Variation in output is expected, so if you need consistency you have to build it in through structure rather than hoping for it. Confidence in the output tells you nothing about accuracy, because fluency and correctness are produced by the same mechanism. And the quality of your input matters more than it would with a search engine, because you are shaping a construction rather than filtering a set of results.
Give it a clear task, a defined role, and real context, and you dramatically improve your odds. Details matter.