CiteGuard: A Multi-Database Cross-Verification Framework for Detecting AI-Hallucinated Citations in Academic Writing.

Executive Summary The rapid integration of Large Language Models into academic writing workflows has accelerated research efficiency, but it has also introduced a critical vulnerability known as citation hallucination. Generative AI systems routinely fabricate references that appear authentic, complete with realistic author names, scholarly journal titles, plausible publication years, and structured Digital Object Identifiers. When these hallucinated references bypass initial editorial checks, they pollute academic literature, waste valuable reviewer time, and undermine scientific integrity. CiteGuard is an automated, open research framework engineered to address this problem directly. The system performs automated, multi-database cross-verification to detect fake or distorted academic citations before publication. By leveraging official metadata registries including Crossref and Semantic Scholar, CiteGuard evaluates the authenticity of cited literature through programmatic API interactions and structural metadata matching. Key Technical Contributions Automated Multi-Database Querying: CiteGuard bypasses single-source limitations by concurrently querying major bibliographical databases, ensuring comprehensive coverage across diverse academic disciplines. Structural Metadata Alignment: The framework parses raw reference strings into distinct components, such as primary authors, publication titles, journal venues, volume numbers, and release dates, checking each field against verified record entries. Advanced Discrepancy Detection: CiteGuard identifies complete hallucinations as well as partial hallucination patterns, such as real authors paired with fabricated paper titles, or legitimate paper titles linked to invalid digital object identifiers. Scalable Verification Architecture: Designed with integration in mind, the system provides a structured mechanism that can be integrated into institutional repositories, journal submission portals, and personal academic workflows. Methodology and System Architecture The framework operates in four primary stages to validate academic references. First, raw reference lists are extracted from input manuscripts and normalized into structured data structures. Second, the system dispatches parallel requests to the Crossref REST API and Semantic Scholar Graph API. These calls query database endpoints using parsed title strings and identifier parameters. Third, CiteGuard analyzes response payloads. If a matching record is returned, the system performs a secondary field-level comparison to verify that author order, publication year, and source journal align within acceptable similarity thresholds. Fourth, the framework generates a detailed integrity score for each reference, flagging unverified citations for human review while providing direct resolution links for verified publications. Practical Applications and Impact CiteGuard serves multiple stakeholders within the research ecosystem. Journal editors and peer reviewers can utilize the framework during the preliminary screening phase to automatically detect non-existent references before assigning reviewers. University libraries and academic departments can implement the system to audit student dissertations and faculty preprints. Furthermore, researchers themselves can run CiteGuard on their draft manuscripts to ensure that AI-assisted writing tools have not introduced erroneous citations into their final bibliographies. Keywords Artificial Intelligence, Citation Hallucination, Large Language Models, Academic Integrity, Automated Verification, Crossref API, Semantic Scholar, Natural Language Processing, Metadata Retrieval. How to Cite Islam, F. (2026). CiteGuard: A Multi-Database Cross-Verification Framework for Detecting AI-Hallucinated Citations in Academic Writing. Zenodo. https://doi.org/10.5281/zenodo.21812883