Generative Artificial Intelligence and Academic Research: How to Uphold Ethical Standards and Scientific Integrity
For scientists, generative artificial intelligence may seem like a way to save time and increase their productivity, but it is not without risks.
Nathalie Guichard, Paris-Saclay University; Agnès Helme-Guizon, Grenoble IAE Graduate School of Management; Grenoble Alpes University (UGA); Anne-Sophie Cases, University of Montpellier; Christelle Aubert Hassouni, ESCP Business School and Jean-François Toti, University of Lille
During the COVID-19 pandemic, refusal to get vaccinated stemmed in part from a lack of confidence in the vaccine’s effectiveness and safety. Climate skeptics (32% in France) continue to deny that human activities are responsible for climate change. These two examples share a common thread: the questioning of scientific findings, leading to “scientoscepticism.”
The integration of artificial intelligence at all levels of scientific production could reinforce this sentiment among the public. Integrity, ethics, and responsibility in research are essential to maintaining public trust. Faced with the upheavals brought about by generative artificial intelligence (GAI) and in the absence of clearly communicated guidelines within the scientific community, researchers in the humanities and social sciences (HSS) are questioning the evolution and ethics of their practices.
In a highly competitive environment where researchers’ careers depend largely on their scientific output, the use of AI-generated content tools may seem like an option—at a supposedly minimal cost (most tools are free and appear easy to use)—to boost their own productivity by leveraging the tools’ ability to synthesize vast datasets and generate content. While this approach may seem appealing, it carries significant risks.
A Two-Pronged Challenge for Researchers
The quality of scientific output depends on apeer-review system, which is usually anonymous. In the publication process, researchers serve either as authors or as reviewers, each with somewhat different challenges.
- According to the authors, IAG can be used at various stages of the research process (literature review, data collection and analysis, writing, etc.), which calls into question the very concept of intellectual property. In the social sciences and humanities, where the research process is often iterative, interpretive, and context-sensitive (due to its subject matter—society and human relationships—and for which qualitative methodologies are frequently employed; for example, a discourse analysis of patients as they adopt an e-health solution), the use of AI can profoundly transform research practices. While it can assist researchers, it also raises a central question: at what point does this support become a delegation of the scientific act itself? However, it is a fundamental principle that, regardless of the extent to which AI is involved in the drafting of a scientific article, only the human author(s) will ultimately be responsible for the content produced. Consequently, there is an urgent need to develop explicit, shared guidelines on how to integrate AI responsibly and to report its use transparently.
- For reviewers, IAG can be integrated into their ownpeer-review process. Indeed, reviewers may be tempted to use it to prepare their reviews (to describe or summarize the article, identify its strengths or weaknesses, or draft certain parts of their report). The IAG may also have been used by the article’s authors, in which case it is up to the reviewers to detect this and assess whether such use complies with the scientific integrity required by the journal.
The use of the Problematic Paper Screener (PPS) tool, created by Guillaume Cabanac, has so far identified 25,000 articles containing “awkward phrasing” that suggests they were written using an algorithm, without any human judgment or proofreading.
However, while many journals have now established guidelines, ethical principles remain vague and practices vary widely, leaving researchers in a state of uncertainty.
Furthermore, it is clear from the discussions and debates within scholarly societies that there is a genuine need for dialogue on these issues, but also a lack of consensus on possible options, given the divergent ethical perspectives.
Toward an Impoverishment of Thought?
A recent study published in the journal Science shows that the adoption of large language models (LLMs) by researchers (across all disciplines) increases their productivity (by more than 89%) and opens up new opportunities (simulating virtual populations when these are difficult to access; rapid synthesis of a body of research articles), but thatit also reduces the quality of their work due to its limited ability to handle complex tasks.
Furthermore, relying on AI for acquiring knowledge and writing carries the risk of standardizing thought and knowledge: its use tends to confine researchers to already well-established fields, thereby reducing scientific exploration “off the beaten path.” Researchers who are aware of these limitations can make the most of AI-powered writing tools by posing precise and specific questions, not settling for the first answer, and systematically verifying the content generated.
The Use of IAG for Article Evaluation: A Major Threat to the Reliability of Research and Its Value to Society?
AI-generated content poses a major problem in that the texts it generates are, for the time being, impossible to distinguish automatically from those produced by humans. Therefore, at this stage, two principles should help us overcome this limitation.
Researchers must be the sole guarantors of the scientific quality of research: researchers are asked to review articles precisely because they are experts in the relevant scientific field. Their analysis must therefore, in theory, be based on their own judgment, without delegation to an IAG. This is, in fact, what is specified in the ethical charters or guidelines of certain academic journal publishers (such asElsevier or Sage), which, at this stage, prohibit reviewers from uploading an article to an AI-based peer review tool. However, if AI-based peer review becomes a support tool, what limits should be set on its use?
The confidentiality of research data is becoming a guarantee of the quality of future research: using a cloud-based AI system to evaluate a paper amounts to taking the risk of exposing confidential and potentially sensitive data to everyone, which is problematic, particularly in the context of evaluating a scientific paper. One option to consider is using a locally installed AI system, not only to preserve the confidentiality of research data but also to avoid feeding the AI with documents that have not been scientifically validated. The credibility of research results and the general public’s trust in scientific discourse are at stake.
The rapid development of AI in the writing and evaluation of scientific articles requires us, on the one hand, to rethink our relationship to knowledge and its construction and, on the other hand, to reflect deeply on how to preserve the ethical foundations of knowledge production and evaluation. Despite some initial relevant initiatives, a collective, societal, interdisciplinary, and international dialogue appears essential to reconcile technological innovation with scientific responsibility.
Nathalie Guichard, University Professor, Paris-Saclay University; Agnès Helme-Guizon, University Professor, Social Marketing, Grenoble IAE Graduate School of Management; University of Grenoble Alpes (UGA); Anne-Sophie Cases, Professor, MRM Laboratory, University of Montpellier, University of Montpellier; Christelle Aubert Hassouni, Lecturer and researcher, specialist in consumer privacy issues, ESCP Business School and Jean-François Toti, Associate Professor of Management Sciences – Marketing, University of Lille
This article is republished from The Conversation under a Creative Commons license. Readthe original article.