How are Australian engineering students actually using generative AI?
A peer-reviewed survey of 131 students at two Australian universities found broad but uneven use: coding was reported by 46.6% of one cohort and 7.1% of the other. Because the cohorts differed in size, year level and location, the study cannot identify which factor caused the gap.
The short version
- What happened
- The study surveyed 131 engineering students: 28 first-year students in one cohort and 103 second-year mechanical-engineering students in another.
- Who it affects
- Teachers need task-specific evidence because students use generative AI differently for coding, tutoring and writing.
- What remains unproven
- The 28-person and 103-person cohorts differed in institution, location, year level and course composition, preventing causal attribution.
Keep in mind
The 28-person and 103-person cohorts differed in institution, location, year level and course composition, preventing causal attribution.
Reported behaviours and benefits were self-reported; the study did not inspect AI logs, assignment quality or learning outcomes.
The sample represented about 30% of the full cohorts and was limited to engineering students at two Australian universities.
Unequal cohort sizes reduce precision for the smaller group and widen uncertainty around its percentages.
The cross-sectional design cannot show how use changes over time or whether training altered behaviour.
Source analysis · 2 linked sources. Primary analysed source date: 12 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: students used AI differently, not uniformly
The strongest message from this Australian study is variation. It surveyed 131 undergraduate engineering students at two universities and found that use depended on the task and cohort. In the smaller cohort, none reported frequent academic use, half reported occasional use and half knew of the tools but did not use them. In the larger cohort, 8.74% reported frequent use, 62.14% occasional use, 28.16% awareness without use and one student reported being unaware.
That is not evidence that one university has solved AI literacy and the other has failed. The cohorts differed in several ways at once: one contained 28 first-year engineering students and the other 103 second-year mechanical-engineering students; they studied at different institutions and in different geographic settings. The comparison reveals what teachers need to ask locally, but it cannot isolate whether year of study, curriculum, institutional communication, location or student composition caused the observed differences.[1][2]
Coding produced the clearest gap between cohorts
Among all participants, 46.60% of the larger cohort reported using generative AI to generate code or help with programming assignments, compared with 7.14% of the smaller cohort. After students who did not use generative AI were removed, the comparison was 50.54% versus 11.11%, and remained statistically significant. Research and writing, and studying or tutoring, were more common in both groups than presentations or visuals.
The coding result is plausible for second-year mechanical-engineering students with more programming exposure, but the design cannot test that explanation. The question also asked whether students had used AI for a purpose, not how often, with which model, or whether the resulting code was correct and understood. A student who once requested a code example and one who routinely delegates assignments could appear in the same category. The result should guide assessment design and teaching questions, not estimate learning gain or misconduct.[1][2]
Students saw speed and ideas—and worried about accuracy
Among students who used generative AI, improved efficiency was selected by 72.22% in the smaller cohort and 61.29% in the larger. Enhanced creativity or idea generation showed a larger difference: 88.89% versus 41.94%. The smaller group also more often selected language help and development of technical skills. These are perceived benefits, collected through checkboxes, rather than measured improvements in creativity, efficiency or technical competence.
Verifying accuracy was the most widely reported challenge—66.67% and 77.42% in the two groups. Difficulty producing relevant content affected 61.11% and 51.61%. Ethical concern was reported by 66.67% of the smaller group and 32.26% of the larger. Only a small minority said they had experienced no challenges. For educators, that combination matters: students already recognise that AI can save time and produce unreliable output, but recognition alone does not show they can diagnose an error in code, a calculation or a cited source.[2]
Policy awareness looked similar, but training evidence is thin
More than 95% of each cohort somewhat or strongly agreed that universities have a role in preparing students to use AI. Median responses for awareness of university policies and ethical use were four on a five-point scale in both groups. The only statistically significant training comparison concerned whether the university provided training on its AI policies: the smaller cohort averaged 4.25 and the larger 3.78, with both medians at four.
These answers do not establish what training students received, whether they attended it, or whether it changed behaviour. The eleven-question survey did not test knowledge of a policy, ask students to resolve an ethical scenario, or score their ability to verify an AI output. A student can agree that training occurred without remembering the permitted uses in a particular assessment. Institutions need comprehension and performance checks alongside satisfaction or awareness surveys.[2]
What this means for teachers and students
Teachers should begin with the actual tasks in a course. A blanket statement that students use ChatGPT hides meaningful differences between coding, tutoring, writing, numerical work and presentations. Each task has different risks and legitimate supports. A programming assessment may need oral explanation, version history and tests that expose whether the student understands generated code. A design exercise may instead examine how students compare ideas, evidence and constraints.
Students need practical verification routines embedded in disciplinary work. For engineering, that includes running tests, checking units and boundary conditions, tracing code, examining assumptions and documenting what the model contributed. Clear rules should distinguish permitted support from work the student must perform independently. The study supports progressive, discipline-specific literacy, but it does not show which teaching method works; that requires comparative learning outcomes rather than a cross-sectional attitude survey.[1][2]
What would change the assessment
A stronger study would sample comparable year groups across several universities, record invitation and response counts, and follow students over time. It would combine self-report with consented usage or assessment evidence and test whether students can detect hallucinated references, unsafe code or flawed calculations. Randomised or well-controlled teaching evaluations could compare different literacy interventions on learning, retention and independent performance.
The present result would become more consequential if similar task-specific patterns recur across disciplines and institutions, or if observed AI use predicts changes in achievement after accounting for prior attainment. It would weaken if verified usage diverges sharply from survey responses. For now, the paper offers a useful map of questions and local variation—not a national estimate of Australian students, proof of improved learning, or evidence that one institution's policy caused better behaviour.[1][2]