Does ChatGPT make up sources? Yes: 7% to 69% of its citations, by study

How often does ChatGPT make up sources? 7% to 69% of citations, depending on the model and the year

ChatGPT makes up sources at a measured rate of 7% to 69% of the citations it produces, depending on which model and which year you look at. The oldest measurement is the worst: Gravel and colleagues asked ChatGPT medical questions in early 2023 and found 41 of its 59 references, 69%, were invented. The newest is the best: Bernstein and colleagues asked browsing-enabled ChatGPT-5 for references behind 70 orthopaedic guideline recommendations in early 2026 and found 7.13% did not exist. No study found zero, and in our own run a DOI that resolved cleanly led to the wrong paper.

The four rows are four different studies with four different tasks, so read the chart as a direction, not a curve. The table below is every study we could open at its source, with the year the model was queried, the sample and the share of citations that did not exist. Medical, legal and general academic references are all here because the failure is the same in each field.

Two things in that table matter more than the headline rates. First, the fabricated share is only half the damage: in Walters and Wilder, "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated. Likewise, 43% of the real (non-fabricated) GPT-3.5 citations but just 24% of the real GPT-4 citations include substantive citation errors." A reference that exists with the wrong year, volume or authors fails the same reviewer. Second, the rate moves with how obscure the topic is. Linardon and colleagues asked GPT-4o for reviews on three disorders and got 6% invented citations on major depression, 28% on binge eating disorder and 29% on body dysmorphic disorder. The narrower your question, the more of the answer is guessed.

Why does ChatGPT make up sources? It writes the shape of a citation, not the record

ChatGPT makes up sources because it produces the most likely next words, and a citation is a highly regular string: surname, initials, year, title, journal, volume, pages, DOI. Every part of that pattern is easy to continue and none of it is looked up. The model that wrote the reference never opened a database, so nothing in the process can tell a real record from a well-formed one.

OpenAI's own researchers say the second half of the mechanism is training. In Why Language Models Hallucinate (Kalai, Nachum, Vempala and Zhang, September 2025), the abstract opens: "Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty." A guessed citation that happens to be right scores; "I do not know" scores nothing; so the model learns to guess. What an AI hallucination is covers that paper in full.

The fabrications are rarely random. In our run below the model gave a real pair of authors who really published in the Journal of Adolescent Health in 2006 on caffeine and adolescents, and then wrote the wrong title, the wrong issue, the wrong pages and a DOI that belongs to a different article. Each field was drawn from the neighbourhood of the truth. That is why a fake citation survives a glance and why the checks in the fourth section test the fields against each other rather than the citation against your memory.

Does web search stop ChatGPT making up sources? It cuts the rate, not to zero

Web search cuts fabricated citations sharply and does not remove them. Browsing-enabled ChatGPT-5 still invented 7.13% of 2,736 references in Bernstein and colleagues' 2026 study, whose abstract closes: "Citation errors and fabricated references persist in ChatGPT-5. Independent verification of generated references therefore remains necessary." Across 53,090 URLs cited by search-augmented models and deep research agents, Rao, Wong and Callison-Burch (April 2026) found 3% to 13% pointed at pages that do not exist, and the agents that cite most hallucinated most. In law, where retrieval is built in, Stanford's Magesh and colleagues (2024) found Lexis+ AI and Thomson Reuters' tools "each hallucinate between 17% and 33% of the time".

We ran the comparison ourselves on 27 September 2026, on GPT-5.6 Luna through Whizi, where web search is a toggle in the composer. One prompt, sent twice, once with search off and once with it on: "List five peer-reviewed journal articles on caffeine and sleep latency in adolescents. For each give authors, year, article title, journal, volume, pages and DOI." Then every DOI went to doi.org and to the Crossref API, and the record that came back was compared field by field with what the model wrote.

Neither run produced five. With search off the model gave two and declined the rest, writing that it did not "want to invent or misstate DOI details"; with search on it gave three with PubMed links attached and said the results did not support two more. The one bad citation came from the search-off run, and it is the instructive kind: the DOI resolves, so a DOI check alone passes it, and only reading the title on the landing page catches it. Two caveats. This ran through Whizi, whose own system prompt is in play, so it is one model on one date and not a rate. And Whizi loses on the thing this page is about: as of September 2026 it has no citation checker, so a link in the answer is still a link you open yourself, and nobody has measured its 280+ models on this task, so the picker cannot tell you which one invents least. What the web search toggle does is give the model live pages to quote, which is where the three checked-out references came from.

How do you check a ChatGPT citation in 60 seconds? Four steps, in this order

You check a ChatGPT citation by testing its fields against a database: search the exact title, resolve the DOI and read the landing page, match volume and pages to the journal's year, then open the author's publication list. Each step catches a failure the one before it misses, and the whole sequence takes about a minute per reference once you have done it twice.

1. Exact title search. Paste the title in double quotes into Google Scholar or the Crossref search box. A real paper appears once, with the same authors. No hit, or a hit with different authors, ends the check: the citation is not usable as written.
2. DOI resolve, then read. Paste the DOI after https://doi.org/. A 404 means the DOI was invented. A page that loads is not a pass yet, so read the title on it. Our Orbeta citation resolved cleanly to a paper about HPV, and the Crossref record showed volume 37, issue 6, 2005, against the 38(3), 2006 the model wrote.
3. Volume and pages against the year. On the journal's site, open the volume the citation names and confirm the year and the page range. Walters and Wilder found 24% of GPT-4's real citations wrong in a substantive field, and this is where the wrong ones show.
4. The author's publication list. Open the first author's Google Scholar profile, ORCID record or department page and find the paper there. Two real authors who never wrote together is a common fabrication, and it is invisible until this step.

The check is faster if the model has to give you the identifiers. Add this line to any request for sources: "Give the DOI and the PubMed ID for every reference, and mark any reference you cannot give a DOI for as unverified." A model with search on will attach the links; a model with search off will either supply DOIs you can resolve or admit it has none, which in our run is what it did. Never ask the model whether its own citation is real: the confirmation is generated the same way the citation was. How to fact check AI answers is the general two-minute method for claims, numbers and quotes; this sequence is the citation-specific part of it.

For a reference you will submit, file or publish, do all four steps. For one you are only reading, step 1 is enough to decide whether to keep reading.

Is ChatGPT a credible source? Not for the citation itself, and the courts have priced that

ChatGPT is not a credible source for a citation, because the citation is a claim about a record and not the record; it is a fast way to find candidate papers whose real records you then open. That distinction has a running bill. Damien Charlotin's database of court decisions in which a party relied on hallucinated material listed 2,086 cases as of 27 September 2026, 1,430 of them in the United States, 828 filed by lawyers and 33 by judges. Those are the filings a court caught and wrote up. Why AI gets things wrong covers the errors that never reach a judge.

So the decision rule: a citation from a chat model is a lead until a database says otherwise, and a lead is worth exactly the minute it takes to check.