How to stop ChatGPT from making up citations
Last updated: July 26, 2026
You cannot fully stop ChatGPT from inventing citations, but you can reduce it: ask only for sources it can retrieve, forbid guessing, and require it to say 'no source found.' Then verify every reference by opening it and confirming it actually supports the sentence, because prompt hygiene lowers the rate but cannot guarantee real citations.
Why ChatGPT invents citations in the first place
ChatGPT generates text by predicting the next token from patterns in its training data. A citation like "(Smith et al., 2019, Journal of X, p. 42)" is just a pattern of tokens, so the model can produce one that looks perfectly formatted without any real paper behind it. It has no built-in ground truth and no default step that checks whether the reference exists.
The consequences are real. In Mata v. Avianca (2023), a lawyer filed a brief citing six ChatGPT-generated court cases that did not exist, and was sanctioned $5,000 (Mata v. Avianca, Inc.). Even specialist tools are not immune: a 2024 Stanford RegLab study found purpose-built legal AI hallucinated on 17% to 33% of queries, including citing real cases that did not support the stated point (Stanford RegLab, 2024).
Prompt hygiene that reduces fabricated citations
Good prompting measurably lowers the fabrication rate. It does not eliminate it. Use these techniques together.
- Give it the sources. Paste the actual documents and instruct: "Cite only from the text I provided; do not use outside sources." This is grounding, and it is the single biggest lever.
- Turn on retrieval. Use a browsing or retrieval-augmented mode so the model quotes real, fetched pages instead of recalling from memory. See retrieval-augmented generation.
- License the non-answer. Add: "If you cannot find a real source for a claim, write 'no source found' instead of inventing one." Models over-cite because they are rewarded for looking complete; give them permission to abstain.
- Ask for verifiable identifiers. Request DOIs, court reporter citations, or direct URLs, which are easier to check and harder to fake convincingly than a bare author-year.
- Separate claim from citation. Ask for the factual claims first, then for sources in a second step, so you can see which claims it cannot actually support.
Why prompt hygiene is not enough
Even with perfect prompts, ChatGPT can still produce a citation that is real but does not back the sentence, a failure the Stanford study calls misgrounding. A real paper, cited accurately, may simply not say what the text claims. That is the core trap: a real citation is not a supported claim.
And you cannot close the gap by asking the model to double-check itself. With no ground truth, it can confidently reaffirm a fabricated or misgrounded reference, exactly as ChatGPT did with the invented cases in Mata v. Avianca. The only reliable fix is to verify against the source itself.
Verify every citation and every claim
After prompting well, do two separate checks on each reference:
- Does the source exist? Open the DOI, URL, or reporter citation. If it will not resolve, treat the reference as fabricated.
- Does it support the sentence? Read the cited passage and confirm it actually states the claim, in context, without cherry-picking.
Mephistopheles runs both checks automatically. Paste an AI answer into /chat, drop a document into /try, or use the browser extension, and it extracts each claim, resolves and reads the cited source, grounds the claim against independent evidence, and returns a per-claim verdict: supported, contradicted, disputed, or unverifiable. On our private benchmark of about 290 hard claims, it catches roughly 88% of factual errors while passing about 98% of true statements clean (about a 2% false-alarm rate). For source-heavy work, see verifying research and academic writing.
What verification cannot do for you
A checker confirms whether a source exists and supports a claim; it does not judge whether the source is the right authority for your context, and it cannot fix a claim that is simply niche enough that no authoritative source exists, where the honest answer is "unverifiable." This is not legal, medical, or financial advice; verified output still needs human judgment before you file or publish.
Frequently asked questions
Why does ChatGPT make up citations?
Because it predicts likely text rather than retrieving facts. A well-formatted citation is just a token pattern the model can generate without a real paper behind it, and there is no default step that checks whether the reference exists or supports the claim.
Can prompting stop ChatGPT from fabricating references?
It can reduce fabrication substantially but not eliminate it. Grounding the model in supplied documents, enabling retrieval, and giving it permission to say 'no source found' all lower the rate. Even then it can cite a real source that does not actually support the sentence, so every citation still needs verifying.
How do I check if a ChatGPT citation is real?
Open the DOI, URL, or reporter citation directly. If it does not resolve, treat it as fabricated. If it does resolve, read the cited passage and confirm it actually states the claim. Mephistopheles automates both steps and returns a per-claim verdict.
Does asking ChatGPT to verify its own citations work?
No. Without ground truth, the model can confidently reaffirm a fabricated or misgrounded reference. In Mata v. Avianca, ChatGPT confirmed cases it had invented. Verify against the actual source instead of asking the model.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.