AI in research: why we need to stop treating every AI-related issue as misconduct

Frontiers | Science news

Science news

Robotics and AI

AI in research: why we need to stop treating every AI-related issue as misconduct<br>AI in research: why we need to stop treating every AI-related issue as misconduct<br>Image: Shutterstock.comAuthor: Simone Ragavooloo, Research Integrity Portfolio Manager<br>AI use in research is now the norm, but the boundaries of acceptable use remain unclear.<br>Our award-winning whitepaper, Unlocking AI’s untapped potential: responsible innovation in research and publishing, found that AI use in research is increasing rapidly year on year, with many feeling a pressure to adopt AI or risk being left behind. Shared norms for using it responsibly, however, are still catching up.<br>At the same time, confidence in whether AI is being used responsibly remains uneven. In our survey, 71% of respondents were concerned about misuse of AI tools, with 53% of researchers stating they had observed what they believe to be ‘AI misuse’ by peers.<br>This point is of critical importance; researchers are adopting AI tools faster than shared norms for responsible use are emerging. If trust in the academic record is to be maintained, the research community needs clearer and more consistent ways to distinguish responsible use from poor practice, misuse and deliberate misconduct.<br>What is AI misuse?<br>The phrase ‘AI misuse’ is now used widely, but often loosely. The term is often used as a catch-all label for everything from honest mistakes, uncritical reliance on AI outputs or missing disclosure, to deliberate fraud.<br>Crucially, what researchers label as misuse is applied inconsistently across the research community. It is used to encompass both deliberate misconduct (e.g. intentional fabrication or deception) and poor or unsafe practices (e.g. uncritical reliance on AI outputs without appropriate validation). This conflation is not harmless: when intentional deception and honest mistakes are treated as the same problem, we lose the ability to respond appropriately, weakening policies and punitive actions, distorting accountability, and undermining effective training.<br>Without a shared definition of AI misuse, views on consequences are becoming polarized<br>The absence of a shared definition is beginning to shape attitudes toward sanctions and accountability. This ambiguity is reflected in our survey findings, where views were highly polarized, ranging from zero-tolerance approaches to more pragmatic calls for transparency:<br>“If an author is found to be using AI, they should be permanently blacklisted from journal publication.”

“Use of AI in writing manuscripts and generating figures should face harsh punishment e.g. ban from publishing in a journal.”

At the same time, others raised concerns about unintentional misuse and the significance of disclosure:<br>“Criminalization of AI should be restricted to undisclosed use. Allow AI tools for publishers, authors, and reviewers with proper guidelines and transparency and flag unintended use.”

“‘Accidental’ plagiarism is a very serious problem, since researchers do not necessarily know the true origins of the ideas they present.”

And more recently, proposals have emerged advocating sanctions such as temporary bans up to one year for issues like hallucinated references, demonstrating the shift toward punitive responses, even where intent may be unclear. While such proposals reflect legitimate concerns about research integrity, they also raise questions about how intent, negligence, and harm should be assessed when determining appropriate responses.<br>In day-to-day reality, AI-related issues (like any form of misconduct) can range from unintended errors due to malpractice to deliberate and harmful deceit with intent.<br>Some existing frameworks already recognize that conduct can sit on a spectrum. For example, work by Sabina Alam at Taylor & Francis where figure 3 presents misconduct on a spectrum - from unintentional errors to deliberate fraud, but in discussions about AI, those distinctions are often flattened into a single category.<br>A practical spectrum of AI use in research<br>Simple scales based on intent alone are not enough. Intent to deceive can be difficult to determine, especially in early-stage or poorly documented cases. But impact alone is not enough either. Some harmful behavior may have limited visible effects at first, while an unintended mistake can still cause serious damage.<br>A robust assessment, therefore, requires consideration of both intent and impact - alongside the context in which the behavior occurred - to distinguish between error, misuse, and misconduct in a consistent and proportionate way.<br>To address this challenge, we propose a simple framework that distinguishes AI-related issues along two dimensions: intent and impact. This creates a more practical basis for assessing whether a case represents responsible use, low-risk misuse, high-risk misuse, or serious...