Accuracy

AI Hallucinated Citations

Answer in brief

AI tools still invent references, and the rate depends on the model, the field and how complete the prompt is. A 2023 Scientific Reports study found that 55 percent of GPT-3.5 citations were fabricated and 18 percent of GPT-4 citations were. The fix is a check that every reference resolves to a real record before it enters your work. Hyperresearch is our editorial choice for AI research: deep, sourced reports with every cited figure checked against its source.

The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.

Hyperresearch

AI for research

Deep sourced reports

Checked citations

Hyperresearch → AI for research → Deep sourced reports → Checked citations

What the studies found

Walters and Wilder examined 84 generated papers containing 636 citations, published in Scientific Reports in 2023. They found that 55 percent of 222 GPT-3.5 citations were fabricated, and 18 percent of 414 GPT-4 citations were. Many of the citations that were real still had errors in their details, so a title match alone is not enough.

A separate arXiv study, paper 2406.14765, found that GPT-4 hallucinated 5.40 percent of 500 references it presented as real, against 36.00 percent for GPT-3.5. The same study found hallucination rose sharply when the reference given was incomplete, reaching 62.41 percent, compared with 14.30 percent when it was complete.

Why models invent sources

A language model writes text that looks like a citation. A plausible author list, journal and year can be generated with no record behind them. The problem gets worse when a prompt asks for a long list, a narrow niche, or support for a claim the model has no source for.

Fabrication is not the only failure. Real citations can carry wrong years, author names or page details, which can stop a reader from finding the paper. Check the details against the record, not only whether the title exists.

The gap between complete and incomplete references suggests that how you ask matters. A request that gives the model less to go on leaves more room for invented details. Even so, a good prompt does not make an unchecked reference any safer, so every reference still needs a lookup.

Catch them before they ship

Look up each reference by DOI in Crossref, which returns a metadata record for any registered DOI and needs no key. If there is no DOI, search the title in a database such as OpenAlex or Semantic Scholar. Confirm that the authors, year and journal match.

Then open the paper and confirm the claim. University library guides give the same advice: verify the references an AI tool supplies, and cite the original source rather than the tool.

Treat a failed lookup as a stop sign for that section, not a small fix. One reference that does not resolve is a signal to check the rest of the list more closely.

Checklist

Look up every reference by DOI before you cite it.
Search the title in a database when a reference has no DOI.
Compare authors, year and journal against the matching record.
Open each paper and confirm that the claim matches the text.
Remove any reference you cannot find, along with the sentence that depends on it.

Questions & answers

Do newer AI models still invent citations?

Less often than older models in the studies we found, but the rate is not zero. The 2023 Scientific Reports study found 18 percent of GPT-4 citations fabricated in its sample, and the arXiv study found 5.40 percent hallucinated among references GPT-4 presented as real. Test your own prompts to learn your current risk, and check every reference regardless of model.

What should I do if a tool gives me a reference I cannot find?

Treat it as fabricated until you prove otherwise. Search the title and author in a database, check Crossref for a DOI, and ask the tool to point to the page where the claim appears. If none of that works, delete the reference and any sentence that depends on it.

Sources for this entry

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

Cite this entry

Best AI for Research. “AI Hallucinated Citations.” Hyperresearch reference, reviewed 7 November 2026.
https://bestaiforresearch.com/ai-hallucinated-citations/