Student AI use can coexist with uncertainty about whether that use is appropriate. In their 2025 study, Manuela Farinosi and Claudio Melchior found that students at the University of Udine valued ChatGPT’s help while expressing concerns about academic integrity, unreliable answers, and dependence. Many wanted clearer guidance from their university.
What interests me here is the distance between knowing how to access a tool and knowing when its assistance belongs in an academic task. The study gives us a useful account of that uncertainty during an early period of adoption. It does not measure whether AI improved or weakened learning.
When permission leaves the learning question unanswered
An instruction such as “use AI responsibly” leaves considerable work for the student. Does asking for an explanation count as acceptable support? What about a summary, an outline, or a paragraph? The answers depend partly on what the assignment asks students to demonstrate.
For an assignment assessing interpretation, an AI-generated explanation may complete work students need to practise themselves. In another task, examining that explanation could be part of the learning activity. Students need to understand the purpose behind a boundary if they are going to make informed decisions beyond a list of permitted tools.
This connects with my discussion of acceptable AI use in assessment. A general institutional position still needs to be translated into expectations for particular tasks.
Student AI use: What the study found
Farinosi and Melchior (2025) surveyed 531 students at the University of Udine in Italy and conducted semi-structured interviews with 60 students. The survey took place during the first semester of the 2023–2024 academic year.
Participants described uses including text summarisation, concept clarification, practice questions, and assistance with academic tasks. The interviews allowed the researchers to examine perceived benefits and concerns in more detail.
Students appreciated convenience and access to explanations, but also worried about accuracy, overreliance, creativity, and opportunities for independent reasoning. Most supported regulation and responsible use, while strongly opposing a blanket prohibition.
Students’ accounts of teaching staff suggested limited practical engagement with the issue. Responses also varied with question wording: more students acknowledged academic use when asked about specific practices than when asked a general question about using ChatGPT for university work.
The authors interpret that discrepancy as possible underreporting linked to ethical uncertainty. I would treat that explanation cautiously: differences in how respondents understood the questions could also matter, and the survey does not establish why their answers differed.
Participation was voluntary, the research involved one university, and its findings describe an early adoption period. Reported concerns about losing skills are perceptions, not measured declines in those skills.
Making expectations useful in teaching
The researchers recommend clear institutional guidance, discipline-sensitive approaches, AI literacy, and continuing student involvement. These recommendations follow from the uncertainty participants described. The study did not test which policy or teaching intervention would work best.
My starting point would be to explain the learning purpose of an assignment before specifying its AI boundaries. If students are learning to construct an argument from sources, we can say which decisions they must make and explain why those decisions matter. That makes the expectation easier to understand than an instruction simply forbidding generated text.
We can then discuss concrete cases. Is it acceptable to ask AI to explain an unfamiliar term? To summarise the assigned reading before reading it? To suggest a counterargument after drafting? Students could compare these examples and explain which parts of the intellectual work remain theirs. This is a suggested teaching adaptation, rather than an activity evaluated in the study.
Verification also needs practice. Ask students to compare an AI explanation with the relevant course text, identify a missing qualification, and decide whether the response is usable. The broader AI Literacy conversation includes this capacity to question assistance and understand its limits. My review of evaluative judgement in the AI age develops the importance of judging quality more directly.
Acknowledgement should describe the assistance received. “I used ChatGPT” tells us little. A brief account of the task, what the tool contributed, and what the student checked provides more useful information. Even then, disclosure alone does not establish understanding; the assignment must provide appropriate evidence of learning.
Finally, invite students to identify unclear boundaries before submission. Their questions may reveal that a policy sounds straightforward to instructors while remaining difficult to apply. Disagreement deserves discussion, including when students worry that a convenient shortcut removes practice they need.
I read this study as a reason to make AI expectations part of teaching. Students were already making decisions about assistance, trust, and academic responsibility. Clear guidance should help them connect those decisions with the learning an assignment is intended to support.
Reference
Farinosi, M., & Melchior, C. (2025). ‘I use ChatGPT, but should I?’ A multi-method analysis of students’ practices and attitudes towards AI in higher education. European Journal of Education, 60(2), e70094. https://doi.org/10.1111/ejed.70094
