A hallucination, in AI terms, is confident falsehood. The model produces a claim, a citation, a quote, or a statistic that is completely fabricated, and it produces it in the same fluent register it uses for everything true.
This is not a bug being patched out. It is a structural property of how these systems work, and it will keep showing up in any text you read or produce with AI assistance.
A language model is trained to produce the most plausible continuation of a text. Its objective is plausibility, not truth. There is no lookup step where it checks a claim against reality.
When a model has seen a pattern a thousand times, such as a citation format or a historical claim structure, it can reproduce the shape of that pattern without having the contents. The output looks correctly formatted because it is correctly formatted. It is simply not true.
Three conditions make this worse:
Fabricated citations are the classic case. A title, an author, a journal, a year, sometimes a DOI, all plausible, none of it existing. This has caused real consequences in legal filings and academic submissions.
Invented quotes are close behind. Models will attribute a reasonable-sounding sentence to a real person with total confidence. The person may hold the general view; they did not say those words.
Fabricated statistics are the most dangerous because they are the most persuasive. A specific percentage attached to a specific year looks like research. It is often arithmetic that no one performed.
Other frequent hallucinations include non-existent books and studies, invented product features, incorrect dates, and biographies that blend two real people.
Hallucinated content is fluent, well-organized, and specific. Specificity is exactly the cue people use to judge credibility, and models produce specificity naturally.
Worse, a fabricated claim usually sits next to three accurate ones. The accurate ones establish trust, and the false one rides along.
You will not catch every hallucination by reading, but you can catch most of them with a few habits.
Treat AI-assisted text as a draft containing leads, not facts.
That last step is worth doing even when nobody asks. It converts a draft into something defensible.
Detection tools cannot tell you whether a claim is true. They can tell you whether a passage looks machine-generated, which is a useful hint: generated passages are where unverified specifics tend to live.
That is why a dedicated hallucination check matters alongside a general AI detector. One asks who wrote this; the other asks whether the claims in it survive scrutiny. Most published hallucinations could have been caught by either check, run before publication.
Assume any specific claim a model produces is unverified until you have checked it. Not because the model is lying, but because it has no mechanism to know the difference.
The writers who use AI well are not the ones with the cleverest prompts. They are the ones who verify. That habit is the entire difference between a useful draft and a liability with good grammar.