The Jagged Edge of Adoption: Comparing GenAI Adoption Between Coding and Non-Coding STEM Graduate Students
Plain Language Summary
Artificial intelligence (AI) is unevenly distributed in higher education; some students lean on it, others barely at all, and we hypothesized that the difference tracks the kinds of work AI happens to be good at. Surveying 371 STEM graduate students at six U.S. institutions, we compared those who use generative AI (GenAI) systems, such as ChatGPT or Claude, for coding against peers who use them for everything except coding. Students who code with GenAI use a wider range of these systems, use them more often, believe more strongly in what the systems can do, and apply them across more of their academic work, with the gap widest on tasks that most resemble coding, such as data analysis and making figures, and narrowest on tasks like editing writing or teaching. We read this as early evidence that graduate students who code are carrying GenAI into the rest of academic life, which means teaching about AI has to account for learners who arrive from very different starting points.
Contribution
Rather than reporting GenAI adoption in higher education as a single rate, we contribute survey evidence that adoption among STEM graduate students is structured by a task's proximity to coding, with effect sizes ordering from "Near" to "Far" tasks in the direction the framing predicts, which positions graduate student coders as mediators of GenAI diffusion rather than simply its heaviest users.
Research Questions
- Do Coders and Non-Coders differ in their confidence in GenAI systems?
- Do Coders and Non-Coders differ in their use of GenAI systems for various educational tasks?
Methods
We analyze complete responses from a Spring 2026 survey of 371 STEM graduate students at six U.S.-based institutions selected for high STEM graduate degree conferral, splitting the sample on a single item, "Do you use GenAI for coding?", into 259 "Coders" (69.8%) and 112 "Non-Coders" (30.2%). The instrument, informed by Diffusion of Innovations (DOI) theory, spanned GenAI perceptions, social dynamics, technologies, identity management, task fit, and institutional conditions; this investigation draws on three sets of items, namely usage across 16 named GenAI systems, from which we computed Breadth, Intensity, and Intensive-margin Intensity composites using only the items each student answered and without imputation, a five-item DOI-inspired confidence composite, and frequency of use across nine academic tasks that we assigned to rough "Near," "Middle," and "Far" buckets by their resemblance to coding. Comparisons use Mann-Whitney U tests with tie-corrected z and rank-biserial correlations for the ordinal items, Welch's t-test with Cohen's d for the confidence composite after a Brown-Forsythe test found the group variances unequal, and Fisher's exact tests with risk differences for a binary "Ever-Used" recoding, with Holm correction applied within families and stratified bootstrap 95% confidence intervals (5,000 resamples, fixed seed). Analyses were conducted in Python 3.13 using the scipy, factor-analyzer, and cvxpy packages.
Key Findings
Coders use a wider range of GenAI systems than Non-Coders (Mdn = 11 vs. 9; rrb = −0.271, pHolm < .001) and use them more frequently (Intensity Mdn = 2.6 vs. 1.9; rrb = −0.416), though the effect diminishes once restricted to the systems each group actually reports using (Intensive-margin Intensity rrb = −0.315). On the DOI-inspired confidence composite, whose internal consistency we judged acceptable (α = 0.739, ω = 0.744, GLB = 0.772), Coders scored significantly higher with a large effect (t = 7.205, df = 152.3, p < .001, d = 0.962), agreeing more readily with items such as "GenAI can complete the tasks needed for my school work." Across the nine tasks, Coders report more frequent use of all but one, navigating graduate school relationships (pHolm = .058), which nonetheless separates the groups on the binary "Ever-Used" recoding, and the effect sizes order as the coding-proximity framing anticipates, with Near items largest (data analysis rrb = −0.417, a 44.5 percentage-point risk difference in ever having used GenAI for it), Middle intermediate, and Far items smallest. The bucket-level contrast confirms a Near-favoring skew (µNear − µFar: rrb = −0.185, pHolm = .005), though its modest size, alongside significant Coder-Non-Coder differences within the Far bucket itself, indicates Non-Coders are also using GenAI for Middle and Far tasks, making the difference one of degree and concentration rather than of kind.
Implications
Because we lacked the demographic, financial, and personality measures Rogers's generalizations invoke, and because the survey did not ask how students learned about GenAI in the first place, we hesitate to label Coders "Early Adopters"; what the findings do support is treating STEM graduate student coders as critical mediators of GenAI diffusion, whose greater confidence and broader task use position them to shape peers' AI practices through informal channels rather than instructional ones. That mediation is consequential for AI education specifically, since an agenda adequate to this unevenness has to accommodate formal, non-formal, and informal learning environments populated by learners whose AI literacies differ and continue to change, rather than assuming a shared baseline that our data suggest does not exist. We read the result as a call for longitudinal, interview-based, and network-analytic follow-up with graduate students negotiating AI in their lives, particularly given a moderate sample (n = 371) and the polarization that an AI-focused survey plausibly selects for.