Detecting AI-Generated Text and Assessing Citation Integrity in Scholarly Communication: A Scoping Review of Methods, Limitations, and Manipulation

Large language models complicate text-origin assessment and increase risks from nonexistent, inaccurate, or strategically used citations. This scoping review maps methods for detecting AI-generated or substantially AI-modified text and for evaluating reference integrity and citation manipulation in scholarly communication. We searched Scopus and Web of Science Core Collection on 31 August 2026, and IEEE Xplore on 2 September 2026. Four thematic search streams yielded 1633 records after filters. Following deduplication, removal of five 2027 records, citation searching, title-and-abstract screening, and full-text assessment, 71 peer-reviewed publications were retained; 69 formed the main synthesis and two provided conceptual background. Detection approaches included statistical signals, conventional machine learning, transformer classifiers, stylometry, watermarking, retrieval, and process-based provenance. Performance varied across generators, domains, languages, text lengths, paraphrases, and mixed human–AI texts, limiting the evidentiary value of a single detector score. Citation-integrity studies addressed fabricated or inaccurate references, claim–source agreement, citation function, coercion, cartels, stacking, self-citation, paper mills, and metadata manipulation. On the basis of this mapped evidence, we propose a layered conceptual audit combining provenance, targeted detection, bibliographic verification, claim-level checking, and network or metadata analysis. The integrated workflow itself has not yet been empirically validated, and consequential decisions should remain subject to documented human review.