How to Stop ChatGPT (and Claude, Gemini, Perplexity) From Making Up Citations

Short answer: ChatGPT invents citations when it is asked to recall sources it cannot look up. To stop it, give it the sources yourself, make it cite only those sources by ID, have it write NO SOURCE when a claim is not supported, and check every DOI and title before you submit. No prompt can guarantee real references, but this workflow makes fakes rare and easy to catch.

That is the idea behind the Sources First Citation Guard: Cite Only by Source ID, NO SOURCE Flags, a Source Manifest, a DOI Verification Table, and a Rescue Audit for AI Drafts prompt on PromptDig. This guide explains why fake references happen, walks through the six step sources first workflow, and shows what to do if you already have a draft full of citations you did not check.

Why ChatGPT makes up citations

A language model writes text that looks like the text it learned from. A reference list has a very regular shape: author names, a year in parentheses, a title, a journal, a volume, pages, and a DOI. When you ask a model to "add five peer reviewed sources," it can produce text in exactly that shape without having looked anything up. The result can look perfect and point to nothing.

Search enabled modes help, because the tool can fetch real pages. They do not remove the problem. A model with search can still attach a real paper to a claim the paper never makes, mix up authors, or cite a page that only mentions the topic in passing. The fix is not a stronger warning in the prompt. The fix is changing where the sources come from.

The sources first workflow (6 steps)

The rule is simple: sources first, sentence second. You find and supply the sources. The AI organizes, drafts, and points to them. You check the result.

Step 1: Find the sources outside the chatbot

Use Google Scholar, PubMed, your library's databases, or your reference manager such as Zotero. Read at least the abstract of each source you plan to use, and the relevant section if a claim depends on it. Pick the sources yourself rather than asking the AI to rank or recommend them.

Step 2: Paste or upload them with IDs

Give each source a short ID: S1, S2, S3. For each one, include the title, authors, year, DOI or URL, and the excerpt or abstract you want the AI to use. If you upload PDFs, still list the IDs and titles in your message so the model has a fixed list.

Step 3: Use a cite only by ID prompt

Tell the model it may cite only the IDs you gave it, and that it must write NO SOURCE when a sentence is not supported by any of them. A short version looks like this:

Use only the sources below, labeled S1 to Sn. After each sentence that
relies on a source, cite it as [S#]. If a sentence is not supported by
any source, write NO SOURCE after it. Do not add any new titles,
authors, years, DOIs, URLs, journals, or page numbers. Do not cite from
memory.

The NO SOURCE tag matters. It gives the model something honest to do when it runs out of support, instead of filling the gap.

Step 4: Ask for a source manifest

After the draft, ask for a table that lists every source the draft uses:

Make a table with these columns: ID, exact title as I gave it, authors,
year, DOI or URL as I gave it, the claim in the draft it supports, and
a quote of no more than 25 words from my excerpt that supports it. If a
claim has no supporting quote, say so.

The manifest makes mismatches visible. If the supporting quote does not actually say what the sentence says, you will see it in one place.

Step 5: Verify each reference in about two minutes

For every source in the manifest:

1. Resolve the DOI. Paste it after https://doi.org/ in your browser. It should open the publisher page for that paper.
2. Match the title and first author. A real DOI can still point to the wrong paper if a digit is off. Make sure the page that opens is the one in your list.
3. Check the claim. Open the paper or abstract and confirm it supports your sentence. This catches the case where the paper is real but says something else.

If a source has no DOI, check the URL, the journal's site, or your library catalog.

Step 6: Fix what fails, without swapping in another unverified source

When a check fails, do not ask the AI to "find a better source." That restarts the problem. Remove the claim, soften it to what your sources do support, or go back to Step 1 and find a real source yourself. Add it to your pack with a new ID and rerun the draft for that paragraph.

Rescue: I already have a draft full of AI citations

This is the situation many people are actually in: a draft that came out of a chatbot with a tidy reference list nobody checked. Do not ask the same chat to "fix the references." Ask it to list them so you can check them yourself:

List every citation in the draft below as a table: the citation as
written, where it appears in the text, the DOI or URL as written, and
what I need to check. Mark every row UNVERIFIED. Do not judge whether
any reference is real and do not fix anything.

Then work down the table with the Step 5 checks. For each reference that fails, look at the sentences that cite it. Remove those claims, rewrite them to match a source you have verified, or find a real source. The RESCUE mode of the Sources First Citation Guard: Cite Only by Source ID, NO SOURCE Flags, a Source Manifest, a DOI Verification Table, and a Rescue Audit for AI Drafts prompt does this listing step plus a replace or remove plan for each failed reference.

Does it matter which AI you use?

The bigger difference is between two kinds of use, not between brands.

• Answering from memory. A plain chat with no tools and no files is drawing on what the model learned in training. Any citation it produces in this mode should be treated as unverified until you check it.
• Answering from documents or search. ChatGPT with search, Claude or Gemini with uploaded files, Perplexity, and notebook style tools that answer from your documents all start from material they can actually see. That reduces invented references, but it does not prevent a real source from being attached to the wrong claim.

So the workflow is the same in every tool: give it the sources, require IDs, ask for a manifest, and check.

Prompts for common research jobs

These PromptDig prompts follow the same no invented sources rule for specific tasks:

• Clean an APA 7 reference list without invented DOIs. The APA 7 Reference List Cleaner from Pasted Citations (No Invented DOIs/Page Numbers) prompt formats what you paste and marks missing fields instead of filling them.
• Build an evidence table from papers you have read. Use the Elicit Evidence Table from Paper Inventory (No Invented DOIs) prompt to organize findings across your own paper list.
• Write a research brief with a source policy. The guide on how to write a Perplexity research brief with a source policy shows how to tell a search tool which sources count.
• Summarize a pasted abstract in plain language. The NIH plain language summary prompt works only from the abstract you paste.
• Scaffold a literature review. The systematic literature review assistant helps with structure, search strings, and screening steps, while you supply the sources.
• Clean BibTeX entries. See how to clean BibTeX without inventing DOIs.

If your problem is invented figures rather than invented sources, the sister guide on how to stop AI making up numbers uses the same idea with a locked numbers pack.

An example, labeled as an illustration

Here is a small run of the DRAFT mode with clearly fictional placeholder sources, so nobody mistakes them for real papers.

The writer pastes two placeholder sources: S1 on overnight sleep and word recall, S2 on naps and a finger tapping task. The claim list has three items: sleep helps word memory, naps help motor skills, and sleep loss lowers exam grades.

The output maps claim one to S1 and claim two to S2, each with the supporting quote. Claim three gets NO SOURCE because nothing in the pack covers exam grades. The draft paragraph cites [S1] and [S2] and leaves a visible NO SOURCE note in place of the exam claim. The reference list shows "[missing: volume]" for S2 instead of making a volume up. The notes suggest search terms for claim three and say to add a real source as S3 first.

The gaps are visible, and nothing in the reference list came from the model's memory.

A short academic integrity note

Check your institution's AI policy before using these tools for graded work. Whatever the policy, you are responsible for every citation in your work. The workflow above is built so you can stand behind each one.

Checklist before you submit

• Every source in the reference list came from your own search, not from a chat.
• Every in text citation points to a source in your list.
• Every DOI resolves at https://doi.org/ to the paper you listed.
• The title and first author on the publisher page match your entry.
• The claim you cite is actually in the source.
• No NO SOURCE tags are left in the text.
• Missing fields such as volume or pages were filled in by you from the publisher page, not by the AI.

FAQ

How do I stop ChatGPT from making up citations?
Give it your sources with IDs, require it to cite only those IDs, have it write NO SOURCE for unsupported claims, and verify every reference yourself.

Why does ChatGPT make up fake references?
It generates text that looks like a reference list. Without access to the sources, it can produce reference shaped text that points to nothing.

Can I trust citations from ChatGPT search or Perplexity?
They are more likely to point to real pages, but a real page can still be cited for a claim it does not make. Check each one.

How do I check if a citation is real?
Paste the DOI after https://doi.org/. Confirm the title and first author match. Then confirm the paper supports your sentence.

What if the DOI is real but the paper says something else?
Treat it as a failed citation. Remove or rewrite the claim, or find a source that supports it.

Is there a prompt that guarantees real sources?
No. A prompt can reduce invented references and make gaps visible. Only checking the sources yourself confirms them.

What should I do with a draft that already has fake citations?
List every reference in a table, check each one, and fix the claims that depend on the ones that fail.

Can I use ChatGPT for a literature review at all?
Yes, for structure, search term ideas, and summarizing papers you supply. Do not use it to find or recall the sources themselves.

Related prompts on PromptDig

Start with the Sources First Citation Guard: Cite Only by Source ID, NO SOURCE Flags, a Source Manifest, a DOI Verification Table, and a Rescue Audit for AI Drafts prompt in DRAFT or RESCUE mode. For more research prompts, see the Research category or Browse more prompts. If you have a citation checking routine that works for your lab or class, Share a prompt.