Open-Source API Pipeline for Detecting Hallucinated References in AI-Assisted Academic Writings
DESIDOC Journal of Library & Information Technology · 13.11.2026 · 46 (6) · 772-778
DOI: 10.14429/djlit.22277 ↗Kurz gefasst
Fachartikel in der Zeitschrift DESIDOC Journal of Library & Information Technology, online erschienen am 13.11.2026. Verfasst von Debojyoti Roy und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 2.5).
Abstract (Originalsprache)
One of the most important dangers of the emergence of Large Language Models (LLMs) in academic writing is the phenomenon of reference hallucinations, which pose serious risks to the bibliographic integrity. Current cloud-based AI detector platforms do not solve the problem effectively due to the use of probabilistic stylistics and the necessity to upload the private intellectual property of a researcher to a third party. This work aims to fulfil three research goals: (i) to create a deterministic and privacy-preserving local citation extraction pipeline; (ii) to validate cross-referencing against international repositories without revealing the manuscript text; (iii) to empirically classify the structural features of fake references. The design for the architecture involves the use of containerised parsing of text to extract bibliographies that are then checked concurrently using the Crossref API. The use of the empirically derived 80% fuzzy string match cut-off allows separation of hallucinations from simple typographic errors. Extracted metadata is subsequently cross-referenced against the Crossref REST API using both DOI-based and bibliographic-search queries. Fuzzy string matching is applied to handle typographic variations in titles with a standard threshold. Based on a benchmark against an AI-generated synthetic corpus of 200 references used to test the pipeline sensitivity, the tool achieved a hallucination rate of 85.0 %, identifying correctly 115 (57.5 %) ghost citations and 55 (27.5 %) frankenstein citations and confirming accurately 30 (15.0 %) references that are truly legit. The pipeline is able to find errors and validate its specific working by being run also on a control corpus of 149 references from the past, solving 107 references only based on text normalisation queries. The proposed framework provides editorial offices, institutional repositories and library professionals with an ethical, deterministic and highly effective tool for pre-submission citation integrity verification in the age of generative AI.
Zitieren
APA
Roy, D., & Roy, P. (2026). Open-Source API Pipeline for Detecting Hallucinated References in AI-Assisted Academic Writings. DESIDOC Journal of Library & Information Technology, 46(6), 772-778. https://doi.org/10.14429/djlit.22277
Vancouver
Roy D, Roy P. Open-Source API Pipeline for Detecting Hallucinated References in AI-Assisted Academic Writings. DESIDOC Journal of Library & Information Technology. 2026;46(6):772-778. doi:10.14429/djlit.22277
BibTeX
@article{Roy2026,
author = {Roy, Debojyoti and Roy, Projes},
title = {Open-Source API Pipeline for Detecting Hallucinated References in AI-Assisted Academic Writings},
journal = {DESIDOC Journal of Library & Information Technology},
year = {2026},
volume = {46},
number = {6},
pages = {772--778},
doi = {10.14429/djlit.22277}
}
EUVIMED