The first time a model invents something at you, it is genuinely disorienting, because it does not sound like a mistake. There is no hedge, no wobble, no change of tone. A fabricated court case arrives in exactly the same voice as a real one.
That is not a figure of speech. In 2023 two New York lawyers were sanctioned in Mata v. Avianca for filing a brief built on cases ChatGPT had invented, with quotations and citations attached. The cases read exactly like law, right up until the judge went looking for them.
That uniformity of tone is the whole problem, and it is not an accident. It follows directly from what the model is doing.
The structural reason
A language model produces text by repeatedly choosing what would plausibly come next. That is the entire operation. It does not look anything up, and it has no store of facts marked as facts. What it has is an extremely detailed sense of how text like this tends to go.
Now consider what happens when you ask for a citation.
The model has seen an enormous number of citations. It knows their shape intimately: author, year, plausible-sounding title, journal, volume, page range. Producing something with that shape is easy and it will always succeed, because the shape is what it learned. Whether the specific paper exists is a different question, and nothing in the process ever asks it.
This is the part worth internalizing: a correct answer and a fabricated one are produced by the same mechanism. One of them happens to match reality. The model cannot tell you which, because it is not doing anything that would let it know.
Which is why the obvious follow-up does not work:
Are you sure?
You are asking the same process to assess itself, and what it produces is a plausible-sounding statement about confidence. That is not a check. It will often cave and "correct" a right answer, or dig in on a wrong one, depending on how you asked rather than on anything about the underlying claim.
Four places it happens most
Fabrication is not evenly distributed. It clusters, and the clusters are predictable enough to build a habit around.
1. Anything with a specific identifier. Citations, DOIs, ISBNs, case numbers, statute sections, API method names, config options, URLs. These have the most learnable shape and the least room to be approximately right. A function that does not exist in a library will still be named exactly what a function that did that job would be named, which is precisely why it looks correct until you run it.
2. Anything recent. A model's training has an end date, and it has no reliable sense of where that line falls or what sits beyond it. It will answer questions about last month with the same composure as questions about the 1800s. Tools that search the web change this, but only for what they actually retrieved. The rest of the answer is still generated the usual way.
3. Anything thin. A small town, a minor figure, a niche library, your company's internal jargon. Where there was little material to learn from, there is more gap to fill, and the filling is invisible. The more obscure your question, the more of the answer is construction.
4. Anything you have signalled the answer to. Ask "what are the three main causes of X" and you will get three causes, whether or not the number three means anything. Ask "why is X better than Y" and you will get reasons, even if the honest answer is that it is not. A question with an answer baked into it is the single most reliable way to be told what you already believe.
The check that actually works
It takes about twenty seconds and it is not a prompt.
Take the most specific, most checkable thing in the answer and verify it exists. Not whether it supports the claim. Whether it exists. Search the exact paper title. Search the case name. Look up the function in the real documentation. Paste the URL into a browser.
Existence is the right test because it is binary, fast, and it is where fabrication fails first. A made-up source usually does not survive being searched for. And it works as a spot check on the whole answer: a model that invented one citation was constructing that whole passage, and the prose around it deserves the same suspicion.
What does not work is asking the model to verify itself, or reading the answer carefully to see whether it feels right. Fabrications are optimized, by construction, to feel right.
Be clear about what this check cannot reach. The third and fourth zones above fabricate without leaving an identifier behind: a confident paragraph about a small company, a third cause invented to round out a list of three. There is nothing to paste into a search box, which is exactly why that filling is invisible. The only defense there is the next section: supply the material yourself and ask for quotes out of it, so there is something to check against.
Reducing it before it starts
Verification is the safety net. These reduce how often you need it.
Supply the material instead of asking from memory. A model working from a document you pasted is doing something much closer to reading, and much further from inventing. This is the single biggest lever, and it is most of what Context Is Everything is about.
Ask for quotes, not summaries. "Quote the passages that address this, exactly, with their headings" is checkable in a way that "summarize what this says about X" is not. You can scan the quotes against the source in seconds.
Give it an explicit way out. "If the document does not cover this, say so and stop" gives it somewhere to stop. Producing a plausible continuation has no natural end point otherwise, so it continues.
Ask open questions. "What causes X?" rather than "what are the three main causes of X?". "Is this approach right for Y?" rather than "why is this approach right for Y?".
The short version
The model is not lying to you and it is not broken. It is producing what an answer looks like, and most of the time that is also a correct answer.
So treat fluency as carrying no information at all, because it does not. Notice when you are in one of the four danger zones. Then take the most specific claim in front of you and check that it exists.
No wording of the question does that job for you.
Next in this track: How to Write Better Prompts covers the shape of the ask, and Context Is Everything covers what to put behind it.