In AI, a hallucination is output that sounds right and is wrong. The model is not lying. It generates the most plausible continuation of the text it has seen, and sometimes plausible and true come apart. The result reads with the same confidence either way.
In code, hallucinations tend to look like this: a function or method that does not exist in the library you are using, a configuration option from a different version, a package name that sounds real but is not, or an API endpoint that was never published. They are more common with newer or less popular libraries, where the model has seen less material.
The good news is that code is easy to check. Run it. A hallucinated function fails immediately with an error saying it is not defined. Tests catch subtler cases. When the model suggests a library, look it up on the package registry before installing it, and confirm the name, the publisher and that it is actively used, because installing a look-alike package can be a security risk.
You can also reduce them up front. Tell the model which versions you are on. Paste the relevant part of the official documentation into the conversation. Ask it to say when it is unsure. Agents that can run code and read real files hallucinate less about your project, because they can look instead of guess.
Engineers have been reading confident, wrong code from each other for decades. Review was invented for exactly this.
In the wild
That `useServerCache` hook doesn’t exist in our version of the framework. Can you check the docs and use the supported API?
