A large language model predicts the next piece of text, over and over, very quickly. That one sentence explains most of what assistants do well — and every one of their characteristic failures. It is worth understanding before you rely on one.
Prediction, not retrieval
The model is not looking up an answer in a database. It is producing text that fits the pattern of the question. When the pattern of a correct answer and the pattern of a plausible-sounding wrong answer are similar, you can get either.
Why speech models are a different shape
A speech model has an objective, checkable target: the words that were actually said. A language model generating prose has no such target. That is why transcription accuracy can be scored with a single number and text generation cannot.
What follows from the mechanism
- Fluency is not knowledge. A well-formed sentence carries no guarantee of truth.
- Context is leverage. Give the model the material and it has less need to invent.
- Confidence is a style, not a signal. Models do not sound less certain when they are wrong.
These guides explain the machinery in plain language: how chat assistants and speech models are built, and why they sometimes state falsehoods with total confidence.