Does ChatGPT make up sources?
Short answer
Yes. ChatGPT and other AI tools make up sources and misattribute real ones. ChatGPT Search named the wrong source for 134 of 200 news excerpts in the Tow Center test (March 2025), and lawyers in Mata v. Avianca were sanctioned for six court decisions ChatGPT invented. Check every reference before you use it.
How often AI tools get sources wrong
Fake citations come in two forms. Some sources do not exist at all: an invented paper, case or URL. Others exist but are the wrong source, or do not say what the AI claims. Both show up in published tests.
The Tow Center for Digital Journalism gave 8 AI search tools 200 news excerpts each and asked them to identify the source of each one. Results within that one test (March 6, 2025):
| Tool | Result in the Tow Center test |
|---|---|
 Perplexity | 37% of answers wrong, the lowest of 8 tools |
| ChatGPT Search | 134 of 200 answers wrong; signalled uncertainty only 15 times and never declined |
| DeepSeek Search | Misattributed 115 of 200 excerpts to the wrong source |
| Grok 3 | 94% of answers wrong; 154 of 200 answers led to error pages |
| All 8 tools | More than 60% of answers wrong |
Two findings matter for anyone who relies on AI for references. First, premium tiers gave more confidently wrong answers than free ones. Second, Grok 3's invented links often sat beside answers that were otherwise correct, so a right answer is no proof of a real link.
Other studies point the same way. Sourcing was the most common problem in the EBU/BBC study of four assistants, affecting 31% of answers; for Gemini it reached 72% (October 21, 2025). And Google AI Overviews gave answers not supported by the sources they cited 56% of the time in the New York Times and Oumi analysis reported by Futurism (April 8, 2026). A citation under an answer does not mean the answer came from it.
Why AI invents citations
A language model learns what a citation looks like far better than which citations exist. Author names, journal titles, years, volume numbers and URL patterns follow strong formats, so the model can produce a reference that looks perfect and points to nothing.
Training also rewards an answer over an admission of doubt. OpenAI's own research found that a model that rarely abstained on factual questions had a 75% error rate on SimpleQA, against 26% for one that abstained 52% of the time (OpenAI, vendor claim, September 2025). Asked for a list of sources, a model that knows only some of them will often fill the list anyway.
Search-connected tools add a second failure: they find a real page but attach it to a claim it does not support, or credit one publisher's reporting to another. The full set of causes is on why does AI hallucinate?
Where fake citations have caused real harm
- Mata v. Avianca, June 22, 2023. Lawyers cited six non-existent decisions generated by ChatGPT and were sanctioned $5,000 (Opinion and Order).
- Wadsworth v. Walmart, February 24, 2025. Motions cited 8 fake cases drafted with Morgan & Morgan's in-house AI; the drafter was fined $3,000 and two others $1,000 each (LawNext).
- Johnson v. Dunn, July 23, 2025. Three Butler Snow attorneys were disqualified and referred to the bar over fabricated ChatGPT citations (EDRM).
- Deloitte Australia, October 6, 2025. A government report with fabricated references and a made-up court quote led to a partial refund of an AU$440,000 contract (The Register).
Courts have addressed AI-hallucinated content in 2,241 decisions in Damien Charlotin's database (checked October 29, 2026). More detail on each case is on real AI hallucination examples.
How to check every AI reference
Run this on every source an AI gives you before you quote it, cite it or hand it in.
Open the link
If it fails to load or lands on a different page, treat the reference as fake until proven otherwise.
Search the exact title
Use Google Scholar, your library catalogue or the publisher's site for papers; a real paper turns up under its title. For a DOI, resolve it at doi.org.
Check the details match
Confirm the authors, year, journal or publisher and page numbers. AI often pairs a real author with a paper they never wrote.
Find the claim inside the source
Search the document for the fact or quote the AI attributes to it. A real source that does not say it is still a false citation.
Look up court cases in a court database
Search the case name and docket number in an official court record or CourtListener, never in another AI tool.
Replace what you cannot confirm
Delete the reference, or find a real source for the claim yourself.
A dedicated checker can help with long drafts. Originality.ai's Fact and AI Hallucination Checker flags inaccurate statements and likely hallucinations, and other options are compared on the best AI fact checkers. None replaces opening the source.
The same habits cover numbers, dates and quotes as well as references: see how to fact-check an AI answer.
Getting fewer fake sources in the first place
You can lower the rate, though never to zero:
- Ask the assistant to say when it cannot find a source. Abstaining cuts errors, per OpenAI's figures above.
- Give it the sources. Paste the documents and ask it to cite only from them. Grounding helps, but even the best model on the Vectara leaderboard adds unsupported content to 1.8% of summaries (October 22, 2026).
- Use a search-connected tool for news, then verify. Perplexity had the lowest wrong-source rate in the Tow test, still 37%. See is Perplexity accurate? and is ChatGPT accurate?
Treat every AI citation as a lead, not a source, until you have opened it and found the claim inside.
Questions
Does ChatGPT make up fake references?
Yes. ChatGPT can invent papers, cases and links that do not exist, and attach real sources to claims they do not support. In Mata v. Avianca, lawyers cited six decisions ChatGPT had generated and were fined $5,000 in June 2023 (Opinion and Order). Check every reference it gives you.
How often does ChatGPT get sources wrong?
In the Tow Center test (March 2025), ChatGPT Search named the wrong source for 134 of 200 news excerpts. It signalled uncertainty only 15 times and never declined to answer. Across all 8 AI search tools tested, more than 60% of answers were wrong.
Which AI is best at citing sources?
Perplexity had the lowest error rate of 8 tools in the Tow Center test, at 37%, and the fewest significant issues of four assistants in the EBU/BBC study, at 30%. That is still wrong often enough that every reference needs checking.
Can I use AI-generated citations in an essay or legal filing?
Only after you have opened each source, confirmed it exists and found the claim inside it. Courts have addressed AI-hallucinated content in 2,241 decisions in Damien Charlotin's database (checked October 29, 2026), and lawyers have been fined and disqualified. The person who signs or submits the work is responsible for every citation in it.
How do I check if a source from ChatGPT is real?
Open the link, then search the exact title in Google Scholar, a library catalogue or the publisher's site. Confirm the authors, year and journal match, and find the quoted claim inside the document. For court cases, search the case name in an official court database. If you cannot confirm it, do not use it.
Sources
- Independent study
- Independent study
- News audit
- Vendor claim
- Benchmark
