Study finds GPT-4o fabricates or errs in over half of mental health research citations
A peer-reviewed study by Deakin University researchers found that GPT-4o fabricates nearly one in five citations in mental health literature reviews, with an additional 45% containing errors. Across the 176 citations tested, 56.2% were either fabricated or inaccurate, with higher error rates for less-studied disorders and specialized topics. The study warns that fabricated citations often point to unrelated real papers, making them difficult to detect without manual verification.
What happened
A study by researchers including Jake Linardon from Deakin University tested GPT-4o's ability to generate citations for mental health literature reviews on major depressive disorder, binge eating disorder, and body dysmorphic disorder. The study found that 19.9% of citations generated by GPT-4o across six simulated literature reviews were entirely fabricated, with an additional 45.4% of real citations containing bibliographic errors such as incorrect DOIs, publication dates, or page numbers.
Of the 176 total citations generated, only 43.8% were both real and accurate, meaning 56.2% were either fabricated or contained errors. Fabrication rates varied significantly by topic: 6% for major depressive disorder, 28% for binge eating disorder, and 29% for body dysmorphic disorder. Among the 35 fabricated citations that included DOIs, 64% linked to real but completely unrelated papers, while 36% were entirely invalid or nonfunctional.
Reported impact
Researchers using GPT-4o for literature reviews risk including fabricated or inaccurate citations in published work, which breaks the chain of verifiability and misleads readers. The hallucinated citations could support false claims since they are often formatted plausibly with real DOIs pointing to unrelated papers, making them difficult to detect without manual verification.
Index analysis
The study demonstrates that GPT-4o's citation accuracy is significantly lower for specialized topics and less-studied subjects. Errors are more prevalent in less familiar areas of mental health research, suggesting the model performs better on widely documented topics.
Governance controls
Researchers and journals are advised to implement rigorous human verification of all AI-generated citations. Academic institutions should strengthen safeguards to prevent fabricated references from entering published work.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.