Normative boundaries of AI in scientific work: Evidence from PhD researchers
Based on a survey of 3,785 international PhD researchers, this study identifies four distinct attitudinal profiles regarding AI in scientific work, revealing that researchers' comfort levels are defined by task-specific boundaries—accepting AI for literature tasks while resisting its use in core intellectual contributions like writing and experimental design—rather than a simple binary of acceptance or rejection.
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
Science has always been a human endeavor, a process where researchers ask questions, gather evidence, and build new understanding through their own minds and hands. But a new tool has entered the laboratory and the library: artificial intelligence. These systems can now write text, analyze numbers, and search through vast amounts of information faster than any person could. This shift raises a fundamental question for the future of discovery: if a machine can do the work, who is doing the thinking? The concern is not just about whether the work gets done, but about how the nature of research changes when humans and machines share the load. If a computer writes a paper or designs an experiment, does the human researcher still deserve the credit? Does the learning process that turns a student into an expert still happen if the machine does the heavy lifting? These are not just technical questions about software; they are questions about the rules and values that govern how science is made.
To understand how the next generation of scientists is navigating this new landscape, researchers Francesco Angelini and Johan Lyrvall looked directly at the people who will be shaping the future of research: doctoral students. They analyzed data from a massive survey of 3,785 PhD students in science, technology, engineering, medicine, and health fields from around the world. Instead of asking a simple yes-or-no question about whether students like using AI, the researchers asked them to rate how comfortable they felt using these tools for five specific parts of their daily work. The tasks ranged from tracking and summarizing scientific literature to writing research articles, collecting and analyzing data, and designing experiments. By looking at how students felt about each specific task, the researchers could see if there was a pattern in their thinking. They found that students did not simply fall into two camps of "lovers" and "haters" of AI. Instead, their attitudes formed four distinct groups, revealing a complex map of where they were willing to let machines help and where they drew the line.
The largest group, representing nearly half of all the students surveyed, showed a clear pattern that the researchers called a "division of labor." These students were quite comfortable using AI for tasks like finding and summarizing existing scientific papers. They saw these activities as support work, where a machine could efficiently sort through information. However, the same students felt much less comfortable using AI for the core intellectual tasks of research: writing the article itself, analyzing the data, or designing the experiments. For this group, the boundary was clear. They were willing to let the machine handle the background work, but they wanted to keep the tasks that required deep thinking, creativity, and final responsibility firmly in human hands. This suggests that for many future scientists, the value of AI is in its ability to assist with information management, not in its ability to replace the human mind in the act of discovery.
Other groups in the study held different views. About one-third of the students were uncomfortable with using AI for almost any research task, a stance the researchers labeled "status quo." They tended to feel uneasy about the technology across the board, though their discomfort was strongest when it came to writing and data analysis. A smaller group, about 16 percent, was the opposite; they were broadly comfortable with using AI for everything, viewing it as a general-purpose tool that could help with any part of the research process. Finally, a small but notable group of 7 percent was "undecided." These students frequently chose "don't know" when asked about their comfort levels, suggesting they were either unsure about the technology, had not used it enough to form an opinion, or were still figuring out where the rules stood. Even among those who were unsure, when they did express a feeling, it was more likely to be positive than negative.
The study also looked at whether factors like the student's field of study, their year in the program, or how often they used AI changed these attitudes. The results showed that students in STEM fields were slightly more likely to be in the skeptical "status quo" group compared to those in medical and health sciences. However, the overall patterns held true regardless of whether a student was in their first year or their final year, or whether they studied full-time. This suggests that the way students view the role of AI is not simply a matter of how long they have had access to the tools, but rather a deeper reflection of how they see the nature of their work. The fact that the "division of labor" group is the largest indicates that the most common emerging norm is not to reject AI entirely, nor to accept it blindly, but to carefully sort which tasks belong to the machine and which must remain human.
These findings have important implications for how science is taught and evaluated. The researchers suggest that the way we think about AI in science needs to move beyond simple rules that either ban or allow its use. Instead, institutions and journals might need to adopt a more nuanced approach that recognizes the difference between using AI to find a reference and using it to write a conclusion. If a student uses AI to summarize a hundred papers, that is a different kind of assistance than using it to generate the core argument of a thesis. The study implies that the future of scientific training will need to focus on helping students understand these boundaries. It is not just about learning to use the tool, but about learning which parts of the research process are essential for developing their own expertise. If a student lets a machine do the writing or the data analysis, they might miss the chance to develop the critical thinking skills that come from doing those tasks themselves.
Ultimately, this research captures a moment of transition. The attitudes of these PhD students are not fixed; they are likely to evolve as the technology improves and as the scientific community establishes new norms. Today, the most common view is a selective one, where AI is welcomed as a helper for information but resisted as a substitute for judgment. This "division of labor" may become the standard way science is practiced, with machines handling the volume of information and humans focusing on the synthesis and creation of new knowledge. As these students graduate and become the leaders, editors, and reviewers of tomorrow, the boundaries they are drawing now will help define what legitimate scientific work looks like in an age of artificial intelligence. The paper does not claim to have solved the debate, but it provides a clear picture of where the line is currently being drawn by the very people who will be crossing it in the years to come.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.