Expectations and Practices around AI Disclosure in CS Research
This paper investigates the misalignment between current AI disclosure policies and researcher expectations in computer science, revealing that while researchers prioritize disclosing AI use for low-involvement tasks like research design, actual disclosure practices frequently over-report less critical uses like writing assistance, prompting recommendations to better align policies with these expectations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern laboratory of computer science, a new kind of assistant has quietly taken up residence. These are generative artificial intelligence tools, software capable of writing sentences, debugging code, and organizing ideas with a fluency that once seemed impossible for machines. Researchers have begun to rely on them to speed up their work, from the initial spark of an idea to the final polish of a manuscript. However, this convenience has sparked a quiet but urgent debate within the scientific community: when a machine helps write a paper, who gets the credit, and how much of that help should be revealed to the reader? The core of the issue is trust. Science relies on the understanding that the work presented is the product of human thought and effort. If a tool does the heavy lifting, the integrity of the discovery could be compromised unless the reader knows exactly what was done by a person and what was done by a machine. Consequently, many top computer science conferences have begun to require authors to include a statement disclosing their use of these tools, a rule intended to maintain transparency in an era of rapid technological change.
A team of researchers at the Indian Institute of Science in Bengaluru set out to investigate whether these new rules are actually working as intended. They wanted to know if the policies written by conference organizers matched what researchers actually thought was necessary, and whether the statements authors were writing in their papers reflected those expectations. To do this, they first looked at the rules themselves. They examined the guidelines of sixty-five major computer science conferences and the four largest professional societies in the field. They found that while most of these venues now have a policy requiring disclosure, the rules are often vague. They tell authors to be honest but rarely specify exactly which tasks require a mention or what details must be included. It is as if a teacher told students to "be careful with your homework" without explaining what "careful" looks like for different assignments.
To fill in the blanks, the researchers asked one hundred and nine computer scientists a simple question: for which parts of the research process is it truly important to admit you used an AI tool? They presented the researchers with a list of twenty-one common tasks, ranging from generating new ideas and designing experiments to writing code and polishing grammar. They also asked them to consider how much human control was involved in each task. The results were revealing. The researchers found that the community generally agrees that disclosure is most critical when AI is used for the heavy intellectual work, such as designing the study, analyzing data, or formulating new hypotheses. In these areas, the human contribution is most vulnerable to being overshadowed by the machine. Conversely, the researchers found that using AI to simply fix typos, rephrase sentences, or format references was seen as far less critical to disclose. Interestingly, the level of human control mattered immensely. When a researcher used an AI tool but kept a close watch, verifying every output and making the final decisions, the need to disclose felt lower. But when the AI was allowed to work with little supervision, the demand for transparency rose sharply.
The team then turned their attention to what was actually happening in the real world. They collected and analyzed nearly fourteen thousand disclosure statements from papers submitted to two major conferences, EMNLP and ICLR. They compared these real-world statements against the expectations they had just gathered from their survey. The gap between what people wanted and what they were getting was startling. While the researchers had found that the most important disclosures involved responsibility and the specific tasks performed, the vast majority of the papers they reviewed failed to include this information. Instead, the most common disclosures were for the very tasks the community deemed least important: editing text and fixing code. In many cases, authors wrote short, generic sentences admitting they used AI for "polishing," but omitted any mention of whether the AI had helped with the core science.
Perhaps the most concerning discovery was the presence of identical, copy-pasted statements appearing in dozens of unrelated papers. One specific paragraph, which listed standard AI uses and disclaimed responsibility, appeared verbatim in ninety-five different submissions to a single conference. This suggested that for many authors, the disclosure had become a box to check rather than a genuine account of their workflow. The researchers noted that while some of these statements were comprehensive, their repetition across unrelated work indicated a performative compliance, where authors followed the letter of the rule without engaging with its spirit. The study also found that authors rarely mentioned which specific AI model they used, even though many readers had expressed a desire to know this detail.
Based on these findings, the researchers propose a shift in how these policies are written. They suggest moving away from broad, one-size-fits-all rules toward a system that categorizes research tasks by how necessary disclosure is. They recommend that conferences create a clear list where certain high-stakes tasks, like designing an experiment, are marked as mandatory for disclosure, while lower-stakes tasks, like formatting references, are optional. To help authors navigate this, they also designed a simple template that authors could fill out. This template would prompt them to list exactly which tasks the AI helped with, how much human oversight was involved, and to explicitly state that the authors take full responsibility for the final work. The goal is not to punish the use of helpful tools, but to ensure that the relationship between the human researcher and the machine is clear, honest, and trustworthy. By aligning the rules with the actual expectations of the scientific community, the hope is that transparency can be maintained without stifling the genuine benefits that these new tools bring to the research process.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.