An AI-driven citation verification system for detecting and flagging fabricated scholarly references in documents

Background

The field of academic publishing and scholarly communication is built on the foundation of verifiable citations, which serve as the backbone for the dissemination and validation of research findings. With the rapid advancement of artificial intelligence, particularly large language models (LLMs), there has been a surge in the automated generation of academic content, including reference lists. However, this has introduced a new challenge: the proliferation of fabricated or “hallucinated” citations—references that appear plausible but do not correspond to actual published works. Such fabricated citations undermine the integrity of the academic record, disrupt the chain of evidence essential for scholarly discourse, and erode trust in scientific literature. As the use of AI-generated content becomes more prevalent, there is a pressing need for robust, automated solutions that can ensure the authenticity of citations in electronic documents.

Current approaches to citation management and verification are insufficient to address this emerging problem. Traditional reference management tools like Zotero, EndNote, and Mendeley rely on user-curated libraries and do not perform external validation of citation authenticity, making them incapable of detecting fabricated references that are manually added or generated by AI. Manual verification, while accurate, is highly labor-intensive and impractical for lengthy documents or large volumes of submissions. Plagiarism and style checkers such as Grammarly and Turnitin focus on formatting and textual similarity, but they do not confirm whether a cited work actually exists. This gap leaves academic institutions, publishers, and researchers vulnerable to the inclusion of non-existent references, highlighting the urgent need for automated, AI-driven systems that can validate the existence and accuracy of scholarly citations at scale.

Technology

This technology is an AI-powered system designed to automatically detect fabricated or “hallucinated” scholarly citations within electronic documents. The process begins with the ingestion of a document (such as a PDF or DOCX), from which the system extracts the full text. Utilizing a large language model (LLM), the system identifies and parses the reference list, isolating each citation entry. For citations containing a Digital Object Identifier (DOI), the LLM performs an external query to retrieve the official publication metadata, specifically the authors and the title. The core of the system is a weighted similarity analysis that compares the parsed citation data with the retrieved metadata. This analysis employs both character-based metrics (such as Levenshtein distance) and semantic similarity measures (like vector embeddings and cosine similarity) to generate a verification score. If the combined, weighted score for author and title similarity falls below a predefined threshold, the citation is flagged as potentially fabricated, and a structured report is generated for the user.

What differentiates this solution is its automated, AI-driven approach to validating the authenticity of scholarly references, addressing a critical gap left by existing tools. Traditional reference managers and plagiarism checkers focus on formatting, user-curated libraries, or textual similarity, but they do not confirm the factual existence of cited works. Manual verification, while accurate, is impractical for lengthy documents. This system uniquely leverages LLMs not only for parsing and identification but also for orchestrating real-time external metadata retrieval and sophisticated, weighted similarity analysis. The use of both character-based and semantic similarity functions allows for robust detection of fabricated references, even in the presence of minor typos or paraphrasing. By automating the verification process and providing structured reports, the technology significantly enhances academic integrity and reliability in scholarly publishing, making it a vital tool for researchers, editors, and institutions.

Benefits

• Automatically detects fabricated or AI-generated false scholarly citations to enhance academic integrity.
• Uses large language models (LLMs) to parse references and retrieve official publication metadata via DOIs.
• Employs weighted similarity analysis combining author and title data to verify citation authenticity.
• Incorporates both character-based and semantic similarity metrics to tolerate minor errors and capture deeper meaning.
• Generates structured reports flagging potentially fabricated citations for user review.
• Automates validation of citations in lengthy documents, reducing the impracticality of manual verification.
• Provides an external validation mechanism beyond traditional reference managers and plagiarism checkers.

Applications

• Academic journal manuscript screening
• University thesis authenticity checks
• Publisher editorial workflow integration
• Grant application citation validation
• Peer review support tools

Additional info

This system extracts text from documents, employing an LLM to parse citations and retrieve official metadata via DOI. It conducts a weighted similarity analysis, comparing parsed author/title data against retrieved metadata using character-based or semantic metrics. Citations falling below a verification threshold are flagged as potentially fabricated, generating a structured report.