Why does AI confidently make things up?
Because it is built to predict the most plausible next token, not to verify the truth. Plausible and true overlap most of the time, but not always.
Language models do not know what they know. They generate the most statistically likely continuation of a piece of text. Most of the time, the most likely continuation is also correct — which is why they feel knowledgeable. But when the question is in a thin patch of training data, or when the most likely-sounding answer happens to be wrong, the model produces fluent nonsense with no internal alarm bells.
Hallucinations are not lies. The model has no concept of lying. It is doing what it was trained to do — being plausible.
Three habits cut hallucinations down to manageable levels.
First, ground the model. Paste the source. Ask it to answer only from the paragraphs you give it. If the answer is not in the paragraphs, ask it to say so.
Second, ask for a confidence read. "How sure are you, and what would you check to verify?" forces the model into a more honest posture.
Third, ask the model to argue against itself. "Critique that answer as if you were peer reviewing it. What might be wrong?" gives you a second pass for free.
The goal is not zero hallucination. The goal is to know which answers need a check.