Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
Normative boundaries of AI in scientific work: Evidence from PhD researchers
Citation. Angelini, F. & Lyrvall, J. (2026). Normative boundaries of AI in scientific work: Evidence from PhD researchers [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2608.25678
Summary. This preprint analyzes task-specific comfort with AI among 3,785 self-selected PhD students in STEM and medical and health sciences who completed Nature’s Graduate Survey 2025. A four-class latent-class model finds that the largest group is selectively comfortable with AI for tracking and summarizing literature but less comfortable with AI for writing, data collection and analysis, and experimental design. The authors interpret these configurations as attitudinal patterns with a normative dimension, not direct evidence of shared legitimacy judgments or actual AI use.
Key claims
- The selected four-class model comprised a 44% “division of labour” profile, a 34% broadly uncomfortable “status quo” profile, a 16% broadly comfortable “all-purpose” profile, and a 7% mostly uncertain “undecided” profile.
- The dominant profile was relatively comfortable with AI for literature tracking and summarization while less comfortable with AI for writing, data collection and analysis, and experiment design, suggesting task-specific boundaries around delegation and scientific responsibility.
- More frequent reported AI use was associated with lower likelihood of membership in the “status quo” and “undecided” profiles relative to the “division of labour” profile; the cross-sectional, self-selected survey cannot establish causal relationships or population prevalence.
- The authors argue that AI governance and disclosure should distinguish the activity, role, human control, and supervision involved rather than treating all AI-assisted research as one homogeneous practice.
Evidence & limitations
- The analysis uses five self-reported comfort items from a May–June 2025 international Nature Graduate Survey sample of 3,785 self-selected PhD students across 107 countries and latent class analysis; it measures attitudes, not observed use, legitimacy judgments, learning, or scientific outcomes.
- As an arXiv preprint, the paper has not been peer reviewed; the authors caution that comfort may reflect familiarity, access, confidence, or perceived reliability as well as normative evaluation, and that the sample is not population-representative.