Generative AI Disclosure Cannot Fully Capture Manuscript Provenance in

Bioengineer.org

Source published: Site published:

Original article · فارسی

A study published in Discover Artificial Intelligence says that requiring researchers to disclose their use of generative artificial intelligence (AI) cannot fully show how these tools helped create an academic paper. The study, written by Minoru Matsui, looks at how publishing rules ask researchers to explain their use of AI so that their work is more open and honest. Authors can usually record clear actions, such as which tools they used, what text they kept, and what checks they carried out. However, descriptions such as improving language or creating ideas are often based on the author's memory after the work is finished. These descriptions cannot fully show how repeated conversations with AI may have changed the paper's arguments, added or removed important details, or changed how its ideas fit together.

The study describes a gap between what authors can report about a tool's use and how strongly readers react to that information. Authors may not be able to explain exactly how ideas developed, but research shows that AI disclosures can strongly affect people's opinions of academic work and the researchers behind it. An earlier study by Chen and Jia used a survey experiment involving 806 people. It found that research described as having no machine assistance received the highest scores for quality, ethics, and trust. Scores were lowest when AI was said to have helped with theoretical work, such as creating hypotheses. Using AI for practical tasks, such as writing analytical code, was more widely accepted. Another study by Nakano and colleagues looked at six types of communication and found that saying AI had helped generally lowered people's views of an author's skill, care, and honesty. Schilke and Reimann also found in 13 experiments that disclosing the use of AI reduced trust between people and in organizations because people saw the work as less legitimate.

To explain how readers react to information about technology used in writing, the paper introduces a model called readerly distraction. It describes three stages: noticing the source of the writing, checking the writing for signs of that source, and forming an opinion. The first stage, called provenance salience, happens when readers notice possible AI involvement. This might happen because of a formal disclosure, an editor's warning, something unusual in the surrounding information, or a reader's personal suspicion about the writing style. This can happen even when no AI tool was actually used. In the second stage, called provenance monitoring, readers examine parts of the text for possible signs of automation. They may look at connecting phrases, references, the quality of evidence, and the author's writing style. Readerly distraction happens when readers spend too much time on these surface details, even though those details may have little to do with the tool's actual role or the real risks it may have created.

The model makes an important distinction between checking where writing came from and judging it negatively because of AI use. Evaluative stigma means forming a negative opinion of an author or paper because it is linked to automated writing. Provenance monitoring means spending attention on checking the writing for signs of technology use. According to the model, a reader can carefully look for signs of AI without losing trust in the paper. Another reader might immediately distrust a paper after seeing an AI acknowledgment, without examining the writing closely. However, readerly distraction can lead to negative judgments when ordinary academic writing habits, such as using opposite ideas in a sentence or repeating linking words like however and nevertheless, are treated as proof of AI writing. This can lead readers to wrongly judge a paper's originality, the author's ability, and the effort involved, even though these writing habits cannot reliably show whether AI was used.

The study also points to a growing problem called source over-attribution. This happens when readers or reviewers try to decide which exact phrases, structures, or ideas came from a human and which came from AI, even though they do not have records made while the writing was taking place. Drawing on research about how people write and remember the sources of ideas, the paper explains that people naturally rebuild memories of their writing process rather than keeping a permanent record of where every idea came from. Studies using typing records, eye tracking, and interviews that help writers remember their thought processes show that visible writing actions need context to be understood. These actions do not automatically reveal what a writer intended. Expecting authors or readers to identify the exact source of every idea, rejected draft, or suggestion about how to organize a paper is therefore an unrealistic demand.

This situation creates an unfair reporting burden, especially for researchers working in academic settings where English is the main language. Academic writing has always involved help from other people. Supervisors, peer reviewers, university mentors, colleagues, professional editors, and translators can all change the structure, flow, and style of a paper. These changes are usually not reported or closely examined after publication. By contrast, help from automated writing tools is singled out through required disclosures and increased checks. As a result, equally clear and polished writing may be seen as professional when a person helped produce it but as a sign that the author has less control when software helped.

AI language models are trained using existing academic writing, so they often produce language that looks like traditional scholarly writing. This makes writing style an unreliable way to decide whether AI was involved. Researchers who cannot afford professional human editing and instead use software may therefore face more criticism of their writing. The author warns that suspicion can also unfairly affect researchers who wrote their papers without any technological help. Their careful, traditional academic language may be mistaken for AI-generated text. Research shows that less experienced researchers and scholars outside the main English-speaking academic centers are quickly adopting automated writing tools. However, the paper explains that unequal levels of scrutiny are mainly caused by differences in how visible the use of these tools is, not simply by differences in language or writing style.

To address these problems, the study rejects two extreme approaches: demanding a complete record of every part of the writing process and treating disclosure as unnecessary. Instead, it supports a reporting system that asks for information in proportion to its importance and risk. The system has three main types of disclosure. Methodological reporting covers AI operations that directly affect data collection, the choice of evidence, literature searches, statistical coding, classification, or the combining of research findings. These uses should be explained in the methods section so readers can understand how the results were produced. Governance reporting explains how authors checked, confirmed, changed, or rejected important AI outputs. This gives readers a record of the steps taken to protect quality. Production-provenance disclosure gives a limited account of general help with drafting, reorganizing, and wording. It recognizes that it is impossible to reconstruct every cause of every part of a finished paper.

The framework suggests four main questions to help decide what should be reported. First, did the AI operation affect the evidence, methods, analysis, or conclusions? Second, could the AI output that was used change the reliability or meaning of the final research? Third, what specific checks were carried out to confirm outputs that could be questioned? Fourth, do any remaining details simply reflect the author's memory of how the writing developed? By focusing on risks to the evidence and research findings, this approach avoids demanding complete chat records, lists of every prompt, or made-up percentages showing how much of a paper came from a machine. Such demands can make peer review harder without helping reviewers understand the ideas or judge whether the methods are sound.

Finally, the paper proposes a standard called accountable judgment. This moves beyond simply stating that authors are responsible for their papers. Current publishing rules already make authors responsible for the accuracy and integrity of their work, but simply accepting responsibility does not prove that they understand or have checked what they publish. Accountable judgment requires authors to show four abilities. They must understand the reasons behind the claims they support and know their limits. They must check important outputs against independent evidence. They must have the authority to accept or reject contributions. They must also be ready to explain and defend their claims when other researchers question them. By making authorship depend on the ability to support and explain research claims during peer review, this approach keeps people meaningfully responsible for their work. It also helps prevent guesses about where writing came from from taking attention away from the actual scientific arguments.

Matsui connects this approach to existing ideas about academic authorship. He notes that guidelines from the International Committee of Medical Journal Editors and the CRediT taxonomy already focus on major contributions and careful oversight. Recent work by Luciano Floridi also argues that authors should be judged by their ability to answer for their work, rather than by trying to trace the technical source of every piece of text. Research by J. Rodrigues describes authorship as involving the ability to select, organize, and support work. In addition, studies by C. Novelli and Floridi warn that handing too much work to automated systems may weaken the very habits of critical thinking that researchers need to develop. Matsui explains that accountable judgment does not replace these ideas or try to provide a complete theory of authorship. Instead, it sets out the conditions under which researchers can show that they genuinely support work created with the help of technology when specific claims or methods are questioned.

This framework also applies to large research projects involving dozens or even hundreds of people who divide the work into specialist tasks. Matsui says it would be unreasonable to expect every coauthor to understand every technical method, check every separate dataset, or repeat every complicated analysis on their own. Instead, accountable judgment should match each person's responsibility to the work they actually oversee. For example, a statistician must be able to explain and defend the statistical models. A researcher who organizes the recruitment of clinical study participants must be able to justify how participants were chosen, how data were collected, and how diagnoses were checked. Working with other researchers or using automated tools becomes a problem only when nobody can be identified as responsible for supporting an important claim or when people sign off on a paper without properly checking its contents.

To test these ideas with real evidence, Matsui suggests several experiments that could compare the readerly distraction model with other possible explanations. Researchers could give readers identical academic papers under different conditions. Some readers would receive no information about assistance. Others would be told generally that automation was used, given details about the specific task AI performed, or told that an equivalent amount of help came from a human editor. Researchers could then measure where readers direct their attention. They could use eye tracking, reading times, rereading patterns, and the subjects covered by reviewers' comments. Separate tests could measure how well readers understand and remember the paper, how much they trust it, how capable they think its author is, and whether they would recommend publishing it.

These experiments would allow researchers to test several predictions that could prove the model wrong. The paper predicts that saying a machine helped with writing will make readers examine text features unrelated to the actual task more closely than they would when the same help came from a human editor, even if the writing itself is identical. It also predicts that checking a paper for signs of AI and losing trust in it are separate processes. This means closer checking does not always have to reduce trust, and negative opinions can form even when there is no measurable change in how closely readers examine the text. Matsui stresses that the model can be disproved by evidence. If disclosures about human and machine assistance lead to the same level of checking, or if changes in readers' attention are fully explained by real differences in research methods, the proposed imbalance would not be supported.

Subject of Research: The limits of what generative artificial intelligence disclosures can tell us and how readers react to them in academic writing

Article Title: The epistemic limits of generative AI disclosure in scholarly authorship

Article References: Matsui, M. (2026). The epistemic limits of generative AI disclosure in scholarly authorship. Discover Artificial Intelligence, 6(1), Article 1916. [https://doi.org/10.1007/s44163-026-02808-6](https://doi.org/10.1007/s44163-026-02808-6)

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02808-6

Keywords: generative AI, academic authorship, AI disclosure, readerly distraction, checking the source of writing, negative judgments, accountable judgment, reporting research methods, reporting oversight, reporting writing assistance, publishing ethics, peer review, fair reporting, academic publishing, writing process, language assistance