Hallucination

A hallucination is when a model produces something fluent and confident that turns out to be false. In coding this shows up constantly: a method that does not exist on the library, an import from the wrong package, a config option someone wishes were real, a plausible-looking function signature that is subtly invented.

Why it happens

It follows directly from what a model is. The model predicts the most plausible next tokens; it does not look anything up and it has no internal "is this true" check. Most of the time plausible and correct line up, so it feels reliable. When they diverge, you get an answer that reads perfectly and is wrong. Crucially, the model has no way to tell the difference, which is why hallucinations come with the same confident tone as correct answers.

What actually reduces it

You cannot make a model stop hallucinating, but you can starve the failure mode:

• Give it ground truth. The single biggest lever. Let the agent read the real file, the real types, the actual docs, so the answer comes from your context instead of the model's memory.
• Prefer tools over recall. An agent that runs the code or greps the repo beats one guessing from training.
• Verify, do not trust. Run it, type-check it, click it. Treat every generated API you have not seen before as unverified until proven.