What Should Graduate Engineering Students Learn About Generative AI? A Critical Review of GAI Competency in Engineering Graduate Education
Plain Language Summary
Since ChatGPT arrived, researchers have produced a steady stream of proposals for what students should know about generative AI — "AI literacy," "AI competency," "AI fluency," and so on — but proposals aimed at doctoral engineering students sit awkwardly with how those students actually learn, which happens less in courses than in labs, advising relationships, and the work of producing new knowledge. We read the 31 engineering education conference papers from 2025 that conceptualize AI competency for graduate engineering students, treating each paper's underlying assumptions about what "competency" is as the thing worth examining rather than adding yet another framework to the pile. Sorting those assumptions revealed five distinct ways of thinking about AI competency, one of which — treating it as a teachable, add-on skill set — runs through nearly every paper, while the question of who can even access these systems is largely left unspoken. Our aim is to map these positions as a compass for program designers and researchers rather than to prescribe a single answer.
Contribution
We contribute a phenomenographic outcome space — five axiomatic clusters and a three-axis representation derived from their co-occurrence — of how the 2025 engineering education conference literature conceptualizes doctoral student "AI competency," and show that the dominant Instrumental/Modular framing, carried even by ostensibly critical work, is misfit for a population that produces knowledge and sets disciplinary norms rather than merely receiving instruction.
Research Questions
- How are engineering education researchers conceptualizing doctoral student "AI competency"?
- Which of those conceptualizations fit the distinctive features of the doctoral engineering context?
Methods
We conducted a critical review with a phenomenographic outcome-space commitment, constructing the corpus through a PRISMA-like screening of roughly 3,000 English-language 2025 conference papers drawn from ASEE PEER, IEEE Xplore, the ACM Digital Library, Engineering Village, and SEFI. Because that yield was intractable for manual screening, we used gpt-oss-120b at low temperature with a prompt iterated against a manually screened validation and test set until it reproduced our inclusion decisions, then manually inspected every model output, arriving at a final corpus of 31 papers. For each paper we reconstructed the argument from problem-space to recommendations, extracted its axiomatic assumptions about social life, epistemology, and the ontology of AI, and synthesized them into roughly seven statements of "what AI competency is" — 221 statements in all — which we then clustered inductively by manual comparison, having abandoned an initial vector-embedding approach because it captured surface semantic similarity rather than phenomenographic difference; statement-level assignments were aggregated to paper-level distributions and co-occurrences. The analysis is interpretive and single-coder, which we treat as a stated limit rather than a reason to claim more than the clusters can bear.