6 prompts that make AI cite its sources

A chatbot will almost always give you an answer, but it will less often tell you where that answer came from. And when it does offer a source, there's a real chance the source doesn't exist.

Asking for citations is still worth doing, not because the model's references are automatically reliable, but because the way it responds to a citation request tells us a lot about how much of the answer was grounded in something real and how much was generated on the fly.

Before running any of these prompts, start your conversation with something along the lines of "If you aren't sure of a source, tell me you don't know rather than guessing." A chatbot that's been given permission to say "I don't know" is less likely to fill the gap with something it made up.

1. Where are you getting this from?

What it does: Asks the model to name the source behind a specific claim in its response.

Why it helps: A chatbot's default is to present information without attribution, which makes everything it says look equally grounded. Asking for the source forces a distinction between claims the model can anchor to something specific and claims it generated from patterns in its training data.

How to use it: Point to a specific claim in the response and ask where the model is getting it from. If it names a source confidently, verify that the source exists and says what the model claims it says. If it hedges or gives a vague attribution like "various studies suggest," that's a signal the claim wasn't drawn from a specific source.

2. Give me enough detail to find that source myself.

What it does: Pushes the model past a vague title or author name and toward full bibliographic information: author, title, publication, date, and ideally a DOI or URL.

Why it helps: A fabricated citation often falls apart at the level of specific detail. A model might generate a plausible author name paired with a plausible journal, but when you ask for the volume number, publication date, and DOI together, the gaps become visible. The more detail you ask for, the easier it becomes to verify or rule out the source.

How to use it: After the model names a source, ask for the complete citation. Then search for the paper or article using those details. If the source is real, you'll find it quickly. If the details don't match anything, the model likely constructed the reference from fragments rather than recalling a specific work.

3. Which parts of this answer are from specific sources, and which are you synthesizing on your own?

What it does: Asks the model to separate the claims it can attribute from the claims it assembled.

Why it helps: Most chatbot responses blend sourced information with the model's own synthesis, and the seams between the two are invisible unless we ask. Getting the model to label which sections are drawn from identifiable sources gives us a map of where to direct our verification effort.

How to use it: After getting a response, ask the model to walk through its answer and indicate which statements come from specific sources and which are its own interpretation or aggregation. Focus your independent checking on whichever category concerns you more for the task at hand.

4. Are you confident that source exists, or are you reconstructing what you think it might say?

What it does: Gives the model explicit permission to admit uncertainty about its own citations.

Why it helps: Chatbots default to confidence. When a model generates a citation, it will present that citation with the same fluency whether it's recalling a real paper or assembling a plausible-sounding one from partial memory. This prompt interrupts that default by making "I'm not sure" an acceptable response, which often produces a more honest assessment of the model's own confidence in the reference.

How to use it: Ask this directly after the model provides a citation. Pay attention to the language of the response. A model that's drawing on a well-known source will usually reaffirm with additional detail. A model that was confabulating will often soften its stance or add qualifiers it didn't offer the first time.

5. Can you find a URL or DOI for that?

What it does: Asks the model to produce a direct, verifiable link to the source.

Why it helps: A URL or DOI is the fastest way to confirm whether a source is real. If the model can produce one that resolves to the correct document, the citation is almost certainly genuine. If it produces a broken link, a link to something unrelated, or declines to provide one, the citation needs independent verification.

How to use it: Ask for the link, then open it yourself. Don't assume a properly formatted URL points to a real page. Models are capable of generating URLs that look structurally correct but lead nowhere. The click is the verification step.

6. If any of your citations are uncertain, flag them for me.

What it does: Asks the model to review its own references and mark the ones it's least confident about.

Why it helps: Rather than treating every citation with equal suspicion, this prompt lets the model triage its own references for us. The flagged citations become our priority for independent checking, and the unflagged ones can be verified afterward with less urgency. It's an imperfect filter, since the model can be wrong about its own confidence, but it's a useful first pass.

How to use it: Add this at the end of a response that includes multiple citations. If the model flags several of its own references as uncertain, that tells us the response was leaning more on synthesis than on source material. If it flags none, verify a few anyway to calibrate how well the model's self-assessment holds up.