PhD Students Aren't Rejecting AI — They're Drawing Task-Level Lines

Only 16% of PhD students say they are comfortable letting AI into most of their research work. But the more revealing number in a new study of 3,785 doctoral researchers is this one: 44% — the single largest group — are not rejecting AI at all. They are drawing a line straight through the middle of their own job, welcoming AI on one side of it and blocking it from the other.

Where does that line fall? Not where the adoption debate usually assumes.

[Fact] The finding comes from a working paper by Francesco Angelini and Johan Lyrvall, posted to arXiv on August 26, 2026, which analyses responses from 3,785 PhD students in STEM and medical and health sciences who took part in Nature's Graduate Survey 2025.

A question most surveys skip

Most workplace AI surveys ask a blunt question: do you use it, and how do you feel about it? The answers produce a familiar story of adopters versus holdouts, and that story shapes everything from journal policy to doctoral training. Angelini and Lyrvall asked something narrower and, it turns out, more useful. They measured comfort with AI on five specific research tasks: writing a research article, collecting and analysing data, designing experiments, tracking the scientific literature, and summarising it — each rated on a six-point scale from very comfortable to very uncomfortable, with a don't-know option.

Then, instead of averaging the answers into a single acceptance score, they ran a latent class analysis — a statistical method that groups people by their whole pattern of responses rather than by any category the researchers chose in advance. Four distinct profiles emerged from the data.

[Fact] The largest, at 44%, is what the authors call a "division of labour" profile. A "status quo" profile — broadly uncomfortable with AI across every task — covers 34%. An "all-purpose" profile, comfortable almost everywhere, accounts for just 16%. The remaining 7% are genuinely undecided, expressing substantial uncertainty rather than a stance.

Note what is missing: a majority on either pole. The two poles together — blanket discomfort plus blanket enthusiasm — cover half the sample. The centre of gravity sits with people who say yes and no at the same time, depending on the task in front of them.

Where the line falls

Inside the division-of-labour profile, the split is sharp. [Fact] 60% are comfortable using AI to summarise the scientific literature, and 48% to track it. But only 33% are comfortable letting AI write a research article, 33% with AI collecting and analysing data, and 36% with AI designing experiments.

Two things stand out in those numbers.

First, the gap. Between summarising literature and writing an article — arguably adjacent language skills — comfort drops by 27 percentage points. That drop happens inside a single profile, inside a single job.

Second, the cluster. [Estimate] The three resisted tasks land within three points of each other (33%, 33%, 36%), while the two accepted tasks sit far above at 48% and 60%. That tightness looks less like five separate opinions and more like one underlying principle applied consistently: AI may handle what surrounds the research, but not what constitutes it. Writing, analysis and experimental design are where authorship and responsibility live — and that, the authors argue, is precisely where resistance concentrates. The paper frames these as normative boundaries: lines concerning delegation, authorship and responsibility, not just preference.

The contrast profiles sharpen the picture. In the all-purpose group, 94% are comfortable with AI summarising literature and 77% even with AI writing the article itself. In the status-quo group, 71% are very uncomfortable with AI writing and 76% with AI touching data collection and analysis.

Exposure moves people; discipline barely does

Who ends up in which profile? [Fact] STEM students are more likely than their medical and health sciences peers to sit in the status-quo camp (coefficient 0.307, p<0.01) — a mildly surprising result, given STEM's reputation for tool enthusiasm. Men are modestly more likely to land in the all-purpose group (0.272, p<0.05). But the variable that dwarfs everything else is behaviour: students who use AI weekly or daily are dramatically less likely to be in the status-quo group (coefficient -1.911, p<0.01).

[Estimate] Compare the magnitudes and the usage coefficient is roughly six times the size of the disciplinary one — a comparison the paper's tables permit but do not spell out. What you do with AI predicts your stance far better than what field you are in.

The obvious reading — familiarity breeds acceptance — deserves a caveat. Causality could just as easily run the other way: people who were already comfortable simply use AI more. The paper's cross-sectional design cannot separate the two, and the authors do not pretend it can.

The limits worth stating

This is a preprint, not yet peer-reviewed, and the sample is self-selected: PhD students who chose to answer Nature's survey may feel more strongly about AI, in either direction, than those who ignored it. The authors also stress a subtler limitation. The survey measures comfort, not judgments of legitimacy. Feeling uneasy about AI writing your paper is not the same as declaring it illegitimate. The four profiles are best read as attitudinal configurations with a normative dimension — not a referendum on what AI should be allowed to do in science.

Why this matters outside the lab

The template these students apply travels well beyond academia. The tasks they protect — writing, analysis, design — are the ones tied to credit and accountability. The tasks they delegate — tracking and summarising what others have written — are the ones where being replaced costs nothing professionally.

That is exactly the split now facing most knowledge workers. Data scientists confront the same question daily: which parts of an analysis are genuinely theirs, and which are plumbing? University professors and lecturers, who both produce research and supervise the doctoral students in this very survey, are currently writing the disclosure rules that the 44% will work under.

[Claim] If 44% of the next generation of researchers already operates on a give-AI-this-but-not-that basis, then policies built around a binary — allow AI or ban it — will misfire for the largest group they govern. Journal disclosure requirements, doctoral training and research evaluation will need to become task-specific, because that is how the people being governed actually think.

For individual researchers, and for knowledge workers watching this from adjacent fields, three practical moves follow from the data:

1. Map your own job the way this study did — not "should I use AI?" but "for which task, specifically?" The 27-point gap shows those are different questions with different answers.
2. Decide explicitly which tasks constitute your intellectual contribution. That boundary, more than any tool choice, will define your professional identity as AI capability grows.
3. Build real fluency in the delegable tasks. The -1.911 coefficient suggests that hands-on use, not abstract debate, is what dissolves blanket discomfort — and gives you an informed basis for wherever you choose to draw your own line.

The line through the middle of the research job is being drawn now, task by task. The people drawing it — today's PhD students — are the ones who will spend the longest careers living with it.

Sources

• Francesco Angelini & Johan Lyrvall (2026), "Normative boundaries of AI in scientific work: Evidence from PhD researchers," arXiv:2608.25678 [econ.GN], v1 submitted August 26, 2026, CC BY 4.0. https://arxiv.org/abs/2608.25678
• Underlying data: Nature's Graduate Survey 2025 (3,785 PhD students in STEM and medical and health sciences), as analysed in the paper above.

This article is an AI-assisted analysis. All figures were verified against the primary source (arXiv:2608.25678v1) before publication. The paper is a preprint and has not yet been peer-reviewed.