What is Appropriate AI Use? Perceptions of Addictive AI Use Among STEM Graduate Student Coders
Plain Language Summary
Graduate students in science and engineering increasingly lean on AI chatbots and coding assistants, such as ChatGPT, Claude, or Cursor, to write the software their coursework and research depend on. Surveying STEM graduate students who code with these systems, we found that nearly two-thirds agree that using generative AI (GenAI) "is addictive," and that the students who agree most strongly are also the ones most likely to worry about becoming dependent, to second-guess what they actually know, and to be careful about mentioning their GenAI use to advisors and peers. We do not claim these students are clinically addicted; what concerns us is that a population already facing elevated rates of anxiety and depression describes its own tool use in these terms while norms about what counts as "appropriate" use remain unsettled. On that basis we argue the pattern warrants a coordinated response, and we offer guidance at the institutional, departmental, course, advisor, and student levels.
Contribution
Rather than advancing a clinical account of "AI addiction," we contribute descriptive survey evidence tying STEM graduate student coders' self-reported perceptions of GenAI addictiveness to dependency worry, self-judgment, and concealment from advisors, and translate those correlational patterns into a five-level policy model spanning institution, department, course, advisor, and student that is built around the apprenticeship structure of graduate education rather than the course-centered assumptions shaping most existing GenAI guidance.
Research Questions
- To what extent do STEM graduate students who use GenAI for coding perceive that use as addictive?
- What is the relationship between perceiving GenAI use as addictive and other constructs such as dependency worry, self-judgment, confidence, and concealment of GenAI use from advisors and peers?
Methods
This position paper draws on a survey administered to STEM graduate students across six top-ranked US universities between Fall 2025 and Spring 2026, focusing on the 259 respondents (69.8%) who reported using GenAI systems to code; of that subgroup, 56% were PhD students, 41.7% master's students, and 2.3% were in accelerated programs. The instrument combined Likert and short-answer items covering GenAI perceptions, social dynamics, identity, dependency, and coding-specific use, defined GenAI for respondents through both a textbook definition and current examples (e.g., ChatGPT, Claude, Midjourney, Cursor, Perplexity), and included the focal item "Using GenAI is addictive" rated on a five-point scale. We examined the subgroup using descriptive statistics and Spearman correlations held to a Bonferroni-corrected threshold (α = 0.05/79 = 0.0006), supplemented by Kruskal-Wallis tests across demographic categories. The analysis is deliberately partial, reporting the correlates that motivated the paper's argument rather than an exhaustive treatment of the survey.
Key Findings
Of the graduate student coder subgroup (N = 259), 64.6% scored ≥ 4 on "Using GenAI is addictive" (M = 3.57, Mdn = 4, skew = −0.72), a distribution leaning toward agreement without differentiating across demographic categories, since the Kruskal-Wallis tests were insignificant. Perceived addictiveness correlated most strongly with worry about becoming dependent on GenAI (r = .446) and with wondering whether one uses GenAI more than peers do (r = .431), followed by GenAI's effect on confidence in what one knows (r = .374) and concern that GenAI use would be apparent in one's work (r = .309). Weaker associations connect the item to judging oneself differently when working with GenAI versus without (r = .283), using GenAI because it does not judge (r = .273), worrying about not learning programming concepts deeply (r = .272), and care about when and how GenAI is mentioned to an advisor (r = .223, the one correlate reported at p < .05 and therefore above the Bonferroni-corrected threshold). Read together, the students who find GenAI most addictive are also those most inclined to conceal that use from the advisors on whom their progress depends.
Implications
Because the concealment runs toward advisors specifically, the pattern threatens the mentorship relationship on which the apprenticeship model of graduate training rests, and with it the processes by which disciplinary knowledge in computing is developed and vetted. We therefore argue for an ecological, feedback-driven policy model in which institutions set definitions broad enough to be reinterpreted, departments and advisors clarify those definitions against students' actual research and coursework tasks, and courses align their GenAI policies both with one another and with their own learning outcomes. These claims rest on students' own sense of "addictive" rather than validated clinical instruments, and on perceptions rather than observed usage, so longitudinal and qualitative follow-up remains necessary to establish whether what we observe reflects genuine behavioral dependency or the taboo currently attached to GenAI use.