[Paper Review] Academ-AI: documenting the undisclosed use of generative artificial intelligence in academic publishing

Alex Glynn|arXiv (Cornell University)|Nov 20, 2024

Artificial Intelligence in Healthcare and Education6 citations

TL;DR

The paper presents Academ-AI, a dataset of 500 suspected undeclared AI usages in scholarly work, and analyzes prevalence by publisher, APCs, DOAJ/SJR metrics, and post-publication corrections.

ABSTRACT

Since generative artificial intelligence (AI) tools such as OpenAI's ChatGPT became widely available, researchers have used them in the writing process. The consensus of the academic publishing community is that such usage must be declared in the published article. Academ-AI documents examples of suspected undeclared AI usage in the academic literature, discernible primarily due to the appearance in research papers of idiosyncratic verbiage characteristic of large language model (LLM)-based chatbots. This analysis of the first 768 examples collected reveals that the problem is widespread, penetrating the journals, conference proceedings, and textbooks of highly respected publishers. Undeclared AI seems to appear in journals with higher citation metrics and higher article processing charges (APCs), precisely those outlets that should theoretically have the resources and expertise to avoid such oversights. An extremely small minority of cases are corrected post publication, and the corrections are often insufficient to rectify the problem. The 768 examples analyzed here likely represent a small fraction of the undeclared AI present in the academic literature, much of which may be undetectable. Publishers must enforce their policies against undeclared AI usage in cases that are detectable; this is the best defense currently available to the academic publishing community against the proliferation of undisclosed AI. This is an updated version of a previous preprint.

Motivation & Objective

• Outline the policy landscape on AI authorship and declaration across major publishers.
• Construct and characterize the Academ-AI repository of 500 suspected undeclared AI usages.
• Assess publisher characteristics (APCs) and impact metrics (SJR) associated with represented journals.
• Examine post-publication responses, including corrections and retractions, to undeclared AI usage.
• Discuss implications for research integrity and enforcement of AI-declaration policies.

Proposed method

• Identify suspected AI-generated text in journal articles and conference papers via phrasing patterns and policy statements.
• Manually curate excerpts and extract metadata; store as Markdown files; manage citations with Zotero.
• Tokenize text with quanteda to build a document feature matrix across eight AI-text features.
• Perform statistical comparisons (Chi-squared, Wilcoxon) and currency-adjusted APC/Citation analyses.
• Compare DOAJ-indexed journals and SJR-indexed journals represented in Academ-AI.
• Describe textual indicators (e.g., first-person usage, knowledge cutoffs, “Certainly, here…”, “regenerate response”).

Experimental results

Research questions

• RQ1 What is the prevalence and distribution of undeclared AI-generated content among the first 500 analyzed documents?
• RQ2 Do journals represented in Academ-AI differ from peers in APCs and citation metrics (SJR, h-index)?
• RQ3 What patterns of publisher involvement and editorial responses (retractions/corrections) exist for undeclared AI usage?
• RQ4 Which textual features most reliably indicate AI-generated text in scholarly articles?
• RQ5 How does representation in DOAJ and SJR relate to the Academ-AI dataset?

Key findings

• The dataset comprises 500 documents: 449 journal articles and 51 conference papers.
• 93.2% of publications are from 2022 or later; 6.8% appear before ChatGPT’s release in 2022.
• Articles appear across 345 journals, with 86% representing only a single article.
• Only 13.1% of journal articles were published in journals from major publishers; 88.2% of conference papers were from major publishers.
• At least 295 articles (65.7%) appeared in journals with some APC; median APC across all was US$150 (IQR 50–1,295).
• Major publishers show higher median APCs (US$3,039, IQR 2,191–3,530) than others (US$80, IQR 34–300; P<0.001).
• 35 articles (7.8%) appeared in DOAJ-indexed journals; 28 of these had APCs (84.8%), higher than the DOAJ average (34.3%; P<0.001).
• Seventy journals in the SJR database published 82 Academ-AI articles (18.3%); represented journals had higher median SJR, h-index, outputs, citations, and citations per document (all P<0.001).
• About 3.0% of examples were corrected post-publication; 2.2% formally corrected and 1.0% stealth-corrected.
• Four of eleven formal corrections identified ChatGPT as the AI tool; other cases cited Grammarly or no AI-declaration in the main article; many corrections did not fully align with policy.