In recent months, concerns have arisen within the scientific community regarding the undisclosed use of AI language models, particularly ChatGPT, in the creation of research papers. A paper published in the journal Physica Scripta in August was flagged when a sharp-eyed computer scientist, Guillaume Cabanac, noticed a peculiar phrase on its third page: ‘Regenerate response.’ This phrase, as it turns out, was a telltale sign of AI assistance, specifically from ChatGPT, a popular AI chatbot known for generating coherent text in response to user queries.
Cabanac shared his discovery on PubPeer, a platform where scientists discuss published research, prompting the authors to admit that they had indeed used ChatGPT in drafting their paper. However, they had not disclose this fact during the paper’s submission process. As a result, the journal decided to retract the paper, citing ethical policy violations.
This case is not isolated. Cabanac has identified over a dozen research articles with similar phrases like ‘Regenerate response’ or ‘As an AI language model, I…’ hidden within their text. Many reputable publishers, including Elsevier and Springer Nature, permit the use of AI language models like ChatGPT as long as authors declare their utilization. However, it’s clear that some authors have either forgotten to remove these obvious AI-generated phrases or deliberately bypassed the declaration requirement.
This issue is believed to be just the tip of the iceberg. As AI language models evolve, so do their telltale signs. In fact, ChatGPT’s ‘Regenerate response’ button was replaced with ‘Regenerate’ in a recent update, making it even harder to detect AI assistance. Cabanac has detected undisclosed ChatGPT use in Elsevier journals, further highlighting the scale of the problem.
The rise of AI language models like ChatGPT presents a broader challenge to scientific integrity. While papers authored partially or entirely by AI are not new, they usually exhibit discernable characteristics that distinguish them from human-written works, such as unusual language patterns or mistranslated phrases. However, as AI tools improve, they become increasingly capable of producing text that is nearly indistinguishable from human authorship, making it challenging for reviewers to identify AI-assisted papers.
This issue extends beyond peer-reviewed journals. Cabanac has also uncovered undisclosed ChatGPT use in conference papers and preprints, which are research manuscripts not yet subjected to peer review. In some instances, authors have confessed to using ChatGPT after their work was questioned on PubPeer.
Moreover, the accessibility and power of AI tools like ChatGPT have raised concerns about potential abuse by paper mills—companies that churn out fake research papers for researchers seeking to inflate their publication records. This could exacerbate the problem of fraudulent research in academia.
The situation also sheds light on the strain on peer reviewers, who are often overwhelmed by the sheer volume of papers submitted for review. AI-generated papers can contain false references, which could serve as red flags for reviewers to investigate further. If a reference does not exist or seems dubious, it may signal the potential use of AI assistance in the manuscript.
In the end, this growing issue underscores the need for greater transparency and ethical awareness in scientific publishing. While AI tools can be valuable aids in research, their use should be declared to maintain the integrity of the scientific process. Researchers, publishers, and peer reviewers alike must adapt to the evolving landscape of AI-assisted research to ensure the continued trustworthiness of scholarly publications.