Fabricated Citations in the Biomedical Literature: Measuring Prevalence, Triaging Evidentiary Impact, and Reducing Response Burden - Responsiveness. This project addresses Focus Area 4 by studying artificial intelligence as both a research-integrity risk and a possible defense; Focus Area 5 by measuring and reducing detection and response burden; and Focus Area 2 by developing and evaluating practical policy and training materials. It advances scientific integrity, administrative-burden reduction, and principles of gold standard science through transparency, reproducibility, and scientific rigor. All project products will be freely available to the Public Health Service-funded research community. Purpose. Generative AI can produce fabricated citations, references to studies that do not exist, which may enter the scientific record without detection. In an audit of 2.5 million biomedical papers, our team identified 4,046 fabricated references and found affected 2026 papers at a rate of roughly 1 in 277. Each detected case can require costly expert review to determine what claim the reference supports and whether the paper remains reliable. This 12-month project will measure the problem, validate a scientific-impact rubric, and develop practical defenses. Activities. Aim 1 will screen approximately 7.5 million to 12 million biomedical papers from 2020 through Q2 2027 using staged bibliographic searches and independent human review. It will estimate prevalence within a defined screenable frame and separately measure detection precision and sensitivity. Aim 2 will use 1,200 re-verified references rated by domain experts to validate a rubric for the role and evidentiary importance of fabricated citations and whether a real source could support the affected claim. The resulting evidence packet will support editorial triage without inferring intent or making a research-misconduct finding. Optional AI assistance will undergo locked evaluation and must abstain when uncertain. Aim 3 will engage editors and Research Integrity Officers to develop a vendor-neutral screening protocol, model policy language that separates editorial action from institutional assessment under 42 C.F.R. Part 93, and brief training. A demonstration will measure detection, false escalations, staff time, and real-world workload at the observed prevalence. Expected outcomes. The project will release a reproducible Fabrication Census; a validated evidentiary-impact rubric, expert-rated validation set, and structured evidence packet; and a demonstrated screening protocol, model policy language, training, and free-core open-source reference implementation. It will report expert agreement, detection performance, whether the evidence packet reduces review time without missing high-impact cases, and prevalence-weighted false escalations and staff time per actionable case. Intended beneficiaries. Journal editors and publishers, institutional Research Integrity Officers, PHS-funded researchers and trainees, and the clinicians, policymakers, and patients who depend on a trustworthy scientific record. Estimated number of people served: 0.
