To check whether AI citations really exist and are correct, you must separate the generation of references from their verification and not trust what a language model reports itself. Audits from 2026 show that large language models produce misleading, fabricated or mislabelled sources in 11.4 to 56.8 percent of cases. Practical steps are: check every reference against an independent source such as a DOI register or catalogue, use multi-model checks where several models must name the same source, deploy specialised detection tools that verify metadata, and record for each claim which real source supports it.
Even automated tools reach only about 88 to 90 percent accuracy on real data, so human review and logging of verification steps remain necessary for critical texts.
The trigger is concrete. On 8 May 2026, CIDRAP (Center for Infectious Disease Research and Policy, University of Minnesota) summarised a large-scale review in The Lancet on fabricated sources in biomedical publications. According to the CIDRAP summary of the Lancet review on fabricated sources, the researchers used an AI-based verification system to check 2.5 million articles and 126 million references and found 4,046 probably fabricated references. According to that summary, the incidence of fabricated citations rose twelvefold between 2023 and 2025, which the authors attribute in part to paper mills, misconduct and uncritical use of generative AI.
For anyone using AI to produce texts with source citations — in law, healthcare, supervision, research or policy — this means that a citation that looks correct is not automatically a real source. In our assessment, the most important shift in 2026 is that checking for misleading AI citations is becoming a separate, designed verification layer rather than an assumption.
What exactly does the 2026 research into fabricated citations show?
Several independent studies point in the same direction. The CIDRAP summary of the review in The Lancet explicitly links the rise to generative AI in writing and reference management, and cites earlier studies that reported 30 to 69 percent fabricated references in LLM output in biomedical contexts.
A synthesis preprint titled Detecting Hallucinated and Suspicious Citations on arXiv compares various major audits, including the Lancet work, and counts, according to the authors, 146,932 hallucinated citations in material from 2025 alone. The authors also refine the definitions: a fabricated citation refers to a work that does not exist, an invalid citation contains incorrect metadata, and a suspicious citation shows patterns that point to fabrication. That distinction matters because each type calls for a different check.
How often do language models fabricate sources and what does that depend on?
The scale comes from a cross-model audit titled How LLMs Cite and Why It Matters, published as an arXiv preprint on reference fabrication in AI-assisted academic writing. The authors examined 69,557 citation instances across ten commercially deployed language models and found hallucination rates between 11.4 and 56.8 percent. The outcome depends heavily on the model, the field and the exact prompt.
From that same study come two usable filters:
- Multi-model consensus: if more than three models cite the same work, the likelihood that the reference really exists rises, according to the authors, to 95.6 percent.
- Repetition within a single prompt: simple classification models based on the bib strings can already catch many fabricated references.
These patterns align with the broader principle of multi-model verification and disagreement as a signal: where models contradict each other about a source, that is a reason to check that reference by hand.
Which tools can detect fabricated AI citations and how well do they work?
Serious detection frameworks now exist, but they are not flawless. Two arXiv preprints are relevant here:
- CiteAudit builds a benchmark of thousands of real and fabricated citations with human labels. A multi-agent verification pipeline reaches up to around 97 percent accuracy on simulated datasets, but falls back to about 90 percent on real citations. Even frontier models show a strong drop in F1 score outside benchmark-like conditions.
- CiteCheck verifies whether a citation matches an existing work and whether the metadata are correct. The tool reaches around 88.9 percent accuracy and 88.7 macro-F1, and thereby outperforms baselines from GPT, Claude and Gemini models, including variants with web search.
The message from both is consistent: automated verification can outperform existing models, but on real data an error margin remains. That is precisely why teams increasingly set up verification as a layered stack for hallucination detection rather than as a single button. More background on the broader subject is available in the topic hub on AI verification and source checking.
How do I build source checking as a separate verification layer into my workflow?
This section is our editorial analysis, based on the studies named above. For high-trust workflows, a verification layer that is separate from text generation appears the most sustainable. A concrete pattern:
- Separate generation and verification. Treat an AI's source list as unconfirmed until it has been checked separately.
- Use multi-model checks for critical claims. Have several independent models name the same reference before you trust it.
- Verify against a real source. Check DOI, title, authors and year against a register or catalogue, or via a CiteCheck-like pipeline.
- Log the claim–source relationship. Record for each important assertion which real source supports it, and where evidence is missing.
- Register the verification steps. Note which check was carried out, with what outcome, so that you can later demonstrate that AI sources were not adopted uncritically.
For legal teams this pattern works well alongside a defensible workflow for AI in legal research, in which citation verification and confidentiality are arranged in advance.
In this landscape we position Vera not as a truth machine, but as a verification layer that can route a task through selected independent models and make verification steps, corrections, disagreements and sources visible for inspection, within the bounds of the workflow. This does not certify that output is correct and does not remove the need for human review of hallucinations, but it can help make visible which AI output was used in which workflow and where uncovered claims remain. The professional final judgement remains with the user.
Sources and references
- Review uncovers rising rate of fake references in published biomedical papers
- How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations
- CiteAudit: You Cited It, But Did You Read It? A Benchmark and Detection Framework for Hallucinated Citations in Scientific Writing
- Detecting Hallucinated and Suspicious Citations
- CiteCheck: Retrieval-Grounded Detection of LLM Citation Hallucinations
Sources: The article relies on the CIDRAP summary of the Lancet review on fabricated sources and on four arXiv preprints: How LLMs Cite and Why It Matters, CiteAudit, CiteCheck and Detecting Hallucinated and Suspicious Citations.