PhD researchers appear to draw different lines around artificial intelligence in scientific work, depending on the task. In a preprint, their reported comfort with AI clustered into four profiles rather than one common attitude. The largest group was relatively comfortable using AI to track and summarise scientific literature, but much less comfortable using it for writing, collecting and analysing data, or designing experiments.
That pattern matters because comfort with a tool is not the same as a judgment that its use is legitimate, ethical or appropriate. The survey measured people's reported comfort, not whether they actually used AI, whether they believed a use was acceptable, or whether it affected scientific outcomes.
Four patterns in the responses
The researchers analysed five survey questions covering AI-assisted writing, data collection and analysis, experiment design, literature tracking, and literature summarising. Each question offered six comfort responses plus "don't know"; none of the five task items had missing answers.
They used latent class analysis, a statistical method that groups respondents according to similar answer patterns. The researchers tested models containing from one to five profiles. The four-profile model had the lowest reported ICLbic value, a measure used to compare model fit: 57,453, compared with 57,713 for the five-profile model. The authors used the four-profile solution for their main interpretation.
The fitted model estimated that 44% belonged to the division-of-labour profile, 34% to the status-quo profile, 16% to the all-purpose profile and 7% to the undecided profile. These are model-estimated shares of the survey's response patterns, not prevalence estimates for all PhD researchers.
The division-of-labour group was the biggest. Its members were comparatively at ease with literature tracking and summarising, while showing substantially less comfort with writing, data work and experiment design. The status-quo group reported discomfort across the activities, with the strongest unease again concentrated on writing, data collection and analysis, and experiment design; its discomfort was milder for the literature tasks.
The all-purpose group was broadly comfortable across the activities, although it too showed stronger comfort with tracking and summarising literature than with writing, data work and experiment design. The undecided group mostly chose "don't know". When respondents in that group expressed a view, it leaned positive, but the authors say the pattern could reflect uncertainty or non-use rather than a clearly defined attitude.
Use was linked to the profiles, but not in a causal way
The survey also examined whether profile membership was associated with research field, PhD stage, time commitment and how often respondents used AI. More frequent AI users were estimated to be less likely to fall into the status-quo group and still less likely to be undecided, compared with the division-of-labour group. Because the data were collected at one point in time, the result cannot show whether using AI changed people's views or whether people with different views chose different levels of use.
There was one reported field difference: STEM students were more likely than medical and health-science students to be in the status-quo group rather than the division-of-labour group. The analysis reported no evidence at the stated 5% threshold for different field composition in the other profiles, or for differences linked to PhD stage or time commitment.
What the survey can—and cannot—answer
The analysis drew on Nature's Graduate Survey 2025, conducted among 3,785 self-selected PhD students in hard-science fields across 107 countries. Analyses involving background variables used 3,722 respondents who had no missing covariate data. The profiles were compared as patterns in the answers rather than as experimental treatment groups.
The study's cross-sectional, self-selected design limits what can be concluded. It does not establish that the profiles are stable over time, that they represent the wider PhD population, or that reported comfort predicts actual AI use for any particular task. It also does not connect the profiles with learning, productivity, skill formation, originality or other scientific outcomes.
The authors interpret the profiles as signs of task-specific boundaries around delegation, authorship and responsibility. They argue that guidance should likewise be tailored to particular research tasks, with implications for governance, doctoral training, disclosure and research evaluation. Those recommendations follow the authors' interpretation of the survey patterns; they are not effects tested by the study.
The paper is an arXiv version 1 preprint dated 26 August 2026. The authors state that the article's data are publicly available through the supplied Figshare dataset link.
Paper data and sources
Original title: Normative boundaries of AI in scientific work: Evidence from PhD researchers
Authors: Francesco Angelini, Johan Lyrvall
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text