If you use an AI tool anywhere in your research workflow, there is a measurable chance it has handed you a reference that does not exist. Not a mistyped page number — a paper that was never written, by authors who never collaborated, in a journal issue that never shipped.
This is no longer a hypothetical risk discussed in AI ethics papers. It is showing up in the published record, and it is accelerating.
What the data actually shows
A study published in The Lancet analysed more than 2 million papers and 97 million citations, identifying roughly 4,000 fabricated references across 2,800 papers. The absolute number is small. The trend is not.
- 2023: roughly 1 in 2,828 papers contained at least one fabricated reference
- 2025: roughly 1 in 458 papers — a sixfold increase
- Early 2026: roughly 1 in 277 papers, across the first seven weeks
That is a more than twelvefold rise in three years. The study’s lead author, Maxim Topaz, described the findings as likely conservative.
Peer review is not catching it reliably. In January 2026, an analysis of 4,841 papers accepted to NeurIPS 2025 — one of the most competitive venues in machine learning — found at least 100 confirmed hallucinated citations across 53 papers, roughly 1% of accepted work. A separate taxonomy of those 100 citations found that 66% were total fabrications: invented wholesale rather than corrupted versions of real sources.
Nature has separately estimated that tens of thousands of publications from 2025 may contain invalid AI-generated references.
Why language models invent references
The mechanism is not mysterious, and understanding it tells you where to look for problems.
Large language models predict likely sequences of text. They are not querying a bibliographic database. When you ask for sources on a topic, the model produces text that matches the shape of an academic citation — plausible author names, a real-sounding journal, a well-formed DOI — because it has seen millions of examples of that shape during training.
This is why fabricated citations are so hard to spot by eye. The formatting is usually flawless. That is the one thing the model is genuinely good at.
The three failure modes worth knowing
Not every bad citation is bad in the same way, and each type is caught by a different check.
1. Total fabrication
The entire reference is invented — title, authors, journal, year, DOI. Nothing exists. This is the most common type and, once you actually check, the most obvious. It survives because it passes a visual scan.
2. Corrupted metadata
The paper is real, but the details are garbled — wrong year, wrong volume, wrong author order, or a DOI that resolves to a different article entirely. The model found something real and then misremembered it.
3. Misattributed claims
The most dangerous category, because every automated check passes. The paper exists, the metadata is correct, the DOI resolves — but the source does not actually support the claim it has been attached to. Only reading the source catches this.
A verification workflow that works
The checks below are drawn from published guidance across several research-integrity sources. They are ordered so the fastest checks eliminate the most references first.
Step 1 — Resolve the DOI
Paste the DOI into doi.org. If it 404s, the citation is fabricated. If it resolves to a different paper than the one cited, the metadata is corrupted. Either way, do not use it as written.
Step 2 — Search the exact title
Put the full title in quotation marks and search Crossref, PubMed, or Google Scholar. A real paper appears in multiple independent indexes. A source that surfaces in exactly one obscure place, and nowhere else, deserves scrutiny before you trust it.
Step 3 — Compare metadata field by field
Check author names, journal, volume, issue, pages, and year against the database record. Watch for near-miss author names — a real researcher in the right field, but not an author on that paper, is a common pattern.
Step 4 — Open the source and confirm the claim
This is the step people skip, and it is the only one that catches misattribution. Read the abstract at minimum. If you are citing a specific finding or statistic, find it in the text. A reference can be entirely real and still be wrong support for your sentence.
Step 5 — Check for retraction
A real, correctly-cited paper may still have been retracted. Standard editorial guidance is to avoid citing retracted work as evidence — cite it only when discussing the retraction itself.
How long this actually takes
Manual verification runs roughly two to three minutes per reference. For a 40-reference paper, that is somewhere between 80 minutes and two hours.
That is a real cost, and it is worth being honest about it rather than pretending verification is free. It is also considerably cheaper than a correction notice, and dramatically cheaper than a retraction.
Automated citation checkers can run the DOI and metadata steps in bulk, which meaningfully reduces the time on Steps 1 through 3. They cannot do Step 4. No tool currently reads your manuscript, reads the source, and judges whether one supports the other.
The underlying problem is not the tool
It is tempting to frame this as an AI problem. It is more accurately an incentive problem that AI has made faster to act on.
As Mohammad Hosseini, who studies research integrity at Northwestern, put it: the presence of hallucinated citations shows there are people unwilling to spend half an hour checking a reference list, which reflects how urgently they need to publish — a symptom of an evaluation model that weights publication counts heavily.
Notably, the audit found fabricated citations concentrated among early-career researchers and small teams — the people under the most pressure and with the least support.
The short version
- Treat every AI-suggested reference as a lead, not a source
- Resolve the DOI; search the exact title in at least two databases
- Compare metadata field by field, not at a glance
- Open the source and confirm it supports your specific claim
- Budget the time — roughly two to three minutes per reference
Use AI for exploration. Use databases for verification. The two are not interchangeable, and the gap between them is where fabricated citations live.
Sources
- Topaz et al., study published in The Lancet on fabricated citations in biomedical literature (reported via STAT News)
- Nature — analysis of hallucinated citations in the scientific literature
- Compound Deception in Elite Peer Review — failure mode taxonomy of 100 fabricated citations at NeurIPS 2025
- ScienceDirect — hallucinations in generative AI and scholarly integrity
This article is research-based rather than the product of hands-on tool testing. See our Methodology page for how we evaluate and what that distinction means.