Why Does AI Invent Statistics and Sources, Even When You Ask for Citations?

Ask an AI writing tool for "three statistics on remote work productivity, with sources," and it will hand you three numbers and three citations without hesitation. The numbers can be wrong and the sources can be real websites that never said what you are told they said. Asking the model to only cite real sources does not reliably fix this, because inventing plausible-looking evidence is not a bug the model can simply switch off.

What is actually happening

A language model is not looking anything up by default. It is predicting what a sentence in this shape usually looks like. When you ask for a statistic with a citation, you are asking it to produce text that has the shape of a supported claim. That can mean a number, a percentage sign, a publication name, a year. The model is very good at that shape. It has read millions of real citations and learned exactly what one looks like.

What it has not done is check whether the number it generated appears in the source it named. Those are two separate steps, and only the first one happens automatically. The second step, verification, is not something the model does unless a real, working tool actually fetches the page and compares text against it. Most consumer chat interfaces do not do this by default, even when they can browse.

This is why the errors are so convincing. A fabricated statistic from a language model rarely looks fabricated. It looks exactly like every other statistic you have ever read, because it was built from the same pattern.

A worked example

Here is an illustrative version of the failure, built to show the pattern rather than to accuse any specific tool.

The prompt: "Give me a statistic on how many B2B buyers research a company before contacting sales, with a source."

A plausible, unverified answer: "73% of B2B buyers complete more than half of their research before ever speaking to a salesperson (Forrester, 2024)."

That sentence reads as completely normal industry content. It also has two separate problems, and they are different kinds of problems.

1. The number: It may be close to a real, widely cited research finding, or it may be a smoothed average of several different real numbers, or it may be invented outright. Without checking, you cannot tell which.
2. The source: "Forrester, 2024" names a real, credible firm. It does not name a specific report, a specific page, or a link you can open. A source you cannot open is not a source. It is a label that makes the sentence feel supported.

A properly supported version does not name a firm from memory. It links to one specific, currently accessible page, quotes the exact figure as that page states it, and states the year the source itself gives, not the year the article happens to be written:

In [Report Name], [Publisher] found that [X]% of B2B buyers said they completed most of their research before contacting a vendor. [Link to the report]

If you cannot fill in every bracket with something you have personally opened and read, the claim is not ready to publish.

Why "only cite real sources" is not enough

Telling the model to only use real, verifiable sources changes its output style. It does not give the model a way to check its own memory against reality. It will still generate a URL that looks exactly like a real one, because it has seen thousands of URLs from that domain and knows the pattern.

A URL that resolves to a real domain and a URL that resolves to the specific page making that specific claim are not the same thing, and only opening the link tells you which one you have.

This is also why asking a second AI model to check the first one's work helps, but does not solve the problem by itself. Agreement between two models that were never shown the source material is not verification. It is two guesses that happen to match.

The claim-check worksheet

Before any AI-assisted paragraph with a number, a study, or a named source goes out, run it through this table. It takes a few minutes per claim, and it is the difference between a defensible article and a liability.

The rule that makes this work is that a claim with no completed row does not pass. Not "pass with a caveat," not "soften the wording." If nobody has opened the source and confirmed the passage, the claim gets cut or rewritten as something you can actually support, such as a first-hand observation instead of a borrowed statistic.

What this means for how you use AI to cite

None of this means AI-assisted writing is unsafe to use at volume. It means the verification step cannot be skipped, and it cannot be delegated to the same tool that produced the draft.

Teams that treat drafting and fact-checking as one step get burned. Teams that treat them as two separate steps, with the second step producing a visible trail, do not.

The practical shift is to keep the source material next to the draft as you write, not in a separate tab you will lose track of by paragraph four. Every claim should be traceable back to the exact passage that supports it, before an editor ever has to ask "where did this come from?."

How Weelkit approach this

Weelkit keeps the research attached to the draft as it is written, so every claim carries the passage and source it came from into review. An editor checking a statistic is looking something up in the same workspace, not reconstructing a research trail from a chat history that has already scrolled away.