Research Article
ChatGPT for academic assignments: opportunities, challenges, and student perspectives in the digital classroom
Synthesis: Kumar, Janani and Shobha (2025) surveyed 304 undergraduate and postgraduate students at two universities in Karnataka, India, on their use of ChatGPT for assignments, then modeled paths from digital literacy, perceived usefulness, and ethical awareness to self-reported performance. Digital literacy predicted most strongly (β = 0.393), perceived usefulness next (β = 0.259), and ethical awareness not at all (β = −0.006, p = 0.924). The authors read that null path as evidence that ethics shapes conduct rather than grades, and argue that institutions must treat ethics instruction and digital-skills training as separate investments. The survey also surfaced a problem the model never tested: students who want ChatGPT's premium features pay for them personally, so access to the better tool tracks ability to pay.
Key Findings
- Digital literacy, not enthusiasm for the tool, carried assignment performance. It was the largest predictor in the model, supporting the study's first hypothesis.
- Perceived usefulness added a smaller but independent effect. Students who judged ChatGPT genuinely helpful reported better outcomes, consistent with the technology acceptance model.
- Ethical awareness did not predict performance. Its path was essentially zero, so the hypothesis failed — a result the authors read as ethics working through long-term judgment rather than marks.
- The three predictors explained about a third of the variance, leaving most of the outcome unexplained.
- The measurement model held, with one fit caveat. Reliability and convergent validity were adequate, but one fit index exceeded the threshold the authors themselves apply.
- Cost, not capability, split the student experience. Many students relied on personal subscriptions or premium features to do the work the survey asked about.
What the survey modeled
The sample was drawn randomly from a population of 1,944,670 students at two universities, with 304 respondents completing a self-administered questionnaire between October 2024 and January 2025. Participants were assured of anonymity and told the questions had no right answers, which the authors credit with reducing social desirability pressure. Items adapted from prior instruments were screened by three experts, then checked empirically: an exploratory factor analysis converged in six iterations, dropping one ethical-awareness item (EA5) for weak loading, before confirmatory factor analysis and an HTMT test.
The scaffolding is conventional: digital literacy from Eshet-Alkalai and Ng, ethical awareness from Kohlberg and Floridi, perceived usefulness from TAM supplemented by UTAUT and self-determination theory. Every construct enters as a direct input to performance, which is what makes the ethics result interpretable — the model asked whether ethical awareness moves grades on its own, not whether it conditions how students use the tool.
The ethics null is the useful result
A near-zero correlation between ethical awareness and performance does not mean ethics education is pointless; it means a broad ethics score does not produce better assignment marks within a semester. The authors argue that ethical sensitivity shapes professional conduct and responsible technology use over time, and may function as a contextual or moderating condition rather than a direct driver.
For curriculum designers that is a scoping argument. If ethics is taught to improve work, this study offers no support; if it is taught so students disclose use, cite sources, and resist the shortcut, the outcome to measure is the behavior. The authors recommend situational work — case-based tasks on bias, misinformation, and misuse — paired with verification routines rather than sold as a performance lever.
The literacy path points the other way: students who could navigate, evaluate, and apply generated material got more from the tool, so the skill being rewarded is judgment about output rather than access to it. The authors push training past orientation toward hands-on evaluating, verifying, and revising, the operations the strongest predictor captures.
The affordability finding the model left out
The discussion adds a cost dimension the model never tested — the part most likely to matter to administrators. Institutional licenses typically cover a baseline tier; the features students find most useful sit behind a personal subscription. Where that holds, students who can pay get a materially different tool, and the performance gap the model attributes to literacy and usefulness may partly reflect what each could afford. The authors call this a reminder that accessibility must accompany innovation, and a caution against reading AI-related achievement gaps as skill gaps.
What this means for practice
- Instructors: Make digital literacy the intervention, since it carries the clearest evidence here (β = 0.393). Build assignments around evaluating, verifying, and revising output, and grade the revision rather than only the final text.
- Instructors: Do not expect ethics content to lift marks. Teach it for the behavior you want — disclosure, citation, refusal of the shortcut — and assess that behavior directly.
- Administrators: Audit which ChatGPT features your license covers before assuming equal access. If premium capability is bought individually, the equity gap is real and invisible in aggregate performance data.
- Designers: Show students how the tool improves outlining, drafting, or feedback; perceived usefulness independently predicted performance (β = 0.259).
- Researchers: Keep the measurement sequence — EFA, CFA, HTMT, then SEM — but replace self-reported performance items with institutional records.
Limitations
- Performance, literacy, and ethics were all self-reported, so the model relates perceptions to perceptions rather than to grades or observed work.
- Collection covered a short window that the authors say cannot capture longer-term learning or behavioral change.
- The sample came from two universities in one Indian state and covered ChatGPT only, limiting generalization to other institutions, disciplines, and tools.
- Ethical awareness was modeled as one broad construct, which may have masked differences between academic honesty, data privacy, and authorship.
- The cost discussion is interpretive: no cost variable was measured, and RMSEA (0.103) exceeded the conventional threshold.
Citation
Kumar R, N., Janani, M., & Shobha, B. G. (2025). ChatGPT for academic assignments: opportunities, challenges, and student perspectives in the digital classroom. Journal of Research in Innovative Teaching & Learning.