"It is a temptation to get it to do the work…" Student Experiences of Navigating the Generative AI Landscape in UK Higher Education: A Cross-Institutional Survey with International Comparison

Synthesis: The StudentXGenAI Project surveyed more than 7,000 students across 7 UK institutions (September–December 2025) on GenAI use in their studies, comparing findings with a companion Australian survey. A significant minority of students conscientiously object to GenAI use, while most users are honest most of the time and try to avoid submitting direct GenAI outputs — yet students still use GenAI throughout the entire learning and assessment process, creating a persistent tension between use and integrity.

Key Findings

1. Conscientious objectors. Despite discourse that "everyone is using it," a significant minority of students deliberately refuse to use GenAI — 32% of UK respondents reported not using it for educational purposes, challenging the universal-usage narrative driven by media headlines and think-tank surveys.
2. Mostly honest but pervasive use. For students who do use GenAI, the majority are honest most of the time and try to avoid submitting direct GenAI outputs, yet they use it for a range of tasks throughout the entire learning and assessment process.
3. The accuracy paradox. Students report low trust in GenAI outputs (51% distrust them) yet high confidence in their own ability to generate desired outputs (76%), suggesting they critically engage with and personalize results rather than blindly outsourcing work.
4. International convergence. The UK and Australian findings are strikingly similar yet nuanced, suggesting shared student norms and tensions across institutions and countries.

Background: challenging the universal-usage discourse

Since the release of ChatGPT in late 2022, media coverage and sector surveys such as the HEPI/Kortext Student Generative AI Survey have pushed a narrative of near-universal student adoption, with reported use climbing toward 95%. The authors note that such reports rest on measurement decisions: the framing of survey questions, hidden non-user data, and the growing embeddedness of GenAI in everyday online activity all shape reported prevalence. Prominent Edtech Platform critics warn of "contextlessness" surrounding AI hype, arguing that Generative AI did not arrive in a vacuum but amplified existing pressures on Pedagogies and Teaching Strategies, Assessment, and the perceived value of a degree.

This paper counters the universal-usage claim with a large-scale, multi-institution design. Between September and December 2025, the StudentXGenAI Project adapted an existing Australian survey instrument for UK audiences and administered it at 7 institutions (4 English, 2 Scottish, 1 Northern Irish), collecting 7,087 responses that reduced to 6,610 after data cleaning. The design deliberately mirrors the companion Australian survey run at four institutions in 2024 (Chung et al., 2026), enabling international comparison and addressing methodological inconsistency that has limited earlier small-sample studies.

Who uses GenAI: demographic patterns

Adoption is far from uniform. The survey categorized 68% of respondents as GenAI users and 32% as non-users for educational purposes. Demographic differences were significant though modest in effect size. The strongest pattern concerned home language and enrollment type, with non-native-speaking international students more likely to be regular (weekly or daily) users, a finding the authors link to the high usefulness of translation and language support tasks. Gender showed the second strongest effect: 59% of males were regular users versus 44% of females, while 67% of those identifying as other genders were non-users. Younger (18–19) and older (40+) students were more likely to be non-users, and students without a registered disability were more likely to be regular users (52%) than disabled students (35%).

These patterns raise Equity and Digital Divide concerns. The finding that disabled students use GenAI less regularly, and that accessibility was rarely cited as a motivation, cuts against the assumption that GenAI automatically serves as an Assistive Technology for all learners. The authors flag that demographic differences in use may compound existing inequalities in access to increasingly powerful, often paid, tools.

Trust and the accuracy paradox

A striking result is the disconnect between student trust in GenAI outputs and confidence in their own skill. While 51% of users expressed negative sentiment toward the accuracy of outputs, 76% were confident in their ability to generate desired outputs. The authors interpret this "accuracy paradox" through disruptive-innovation theory: the inferiority of primary functionality (factual and citation accuracy, mathematical rigor) is overridden by secondary functionality such as convenience, conversational accessibility, and 24/7 availability.

Crucially, students report learning GenAI mostly through trial and error (95%), internet searches (92.5%), and using GenAI itself (88.7%), while university-provided workshops and online resources were the least popular and least helpful. The authors see this as evidence that students are not passively consuming AI outputs but actively calibrating their use — yet it also signals a gap in formal AI Literacy and institutional support, with students developing their own ethical compass in isolation from university policy.

Use across the assessment lifecycle

Students use GenAI for a wide range of academic tasks spanning the whole learning and assessment process. The most useful tasks were summarizing, step-by-step instructions, and generating ideas or brainstorming, while the most popular tasks in the UK were finding information and conducting research (89% of users). The authors note that this overlaps directly with learning and assessment activities — brainstorming, structuring work, editing and improving writing — meaning GenAI's influence is present throughout, not merely in producing a finished assignment.

The paper argues this is exactly why GenAI presents a "wicked problem" for institutions, and why it differs from prior academic-integrity moral panics. Unlike copy-paste plagiarism or contract cheating, GenAI can enhance learning rather than simply bypass it. The authors invoke the concept of postplagiarism (after Eaton, 2023) to describe an era where AI is a normal part of teaching, learning, and daily interaction, shifting the focus of assurance from detecting misconduct in the end product to whether students have genuinely attained learning outcomes.

Academic integrity: the honest majority and a minority of Moriarties

The survey asked students whether they use GenAI for part or all of an assessment when not permitted, framing "cheating" around breaking rules or expectations. On a generous reading, 68% of all respondents never use GenAI when not permitted, and a further 21% only use it sometimes — supporting the authors' "Academic Integrity Maxim" that the majority of students are honest most of the time. The habitual rule-breakers who use GenAI "most of the time" or "always" when not permitted form just 5.2% of respondents (7.3% of GenAI users), a "minority of Moriarties" consistent with long-standing estimates that 10–15% of students engage in contract cheating.

A less charitable reading is that roughly 32% of students have broken university policy on GenAI use, and Academic Integrity concerns are compounded by known under-reporting in self-reports. The authors emphasize that much apparent "misuse" may instead be students testing capabilities by trial and error, or justifiably using GenAI to learn when a lecturer's preference or institutional policy forbids it. This complicates simple integrity narratives and points toward the need for nuanced policy rather than blanket prohibition.

Motivations and discouragements: an ethical fast lane

Two themes dominate motivations and discouragements: efficiency and ethics. The top motivations for use were all efficiency-related — 65% said GenAI makes things faster, 57% cited getting unstuck and overcoming blocks, and 55% cited making things easier. Facilitation of work mattered more than improving grades, which 29% of users were not at all motivated by. Meanwhile, the strongest discouragements for users were inaccuracy of outputs (68%), the risk of breaking university rules (55%), and the desire to do the work themselves (51%). Non-users were especially discouraged by wanting to do the work themselves (86%), inaccuracy (81%), and broader ethical and environmental concerns.

The authors interpret this as students seeking an "ethical fast lane rather than an unethical bypass" of learning. The tension between the efficiency motive and ethical discouragements — inaccuracy, rule-breaking, Ethics, environment, and data Privacy — reveals a student body actively trying to regulate its own use. This resonates with calls to protect learner agency and to develop students' capacity to judge whether a given use is helping or hindering learning, rather than leaving them to navigate the gray area of acceptable use alone.

What this means for practice

• Instructors. Design assessment around GenAI being used across the whole lifecycle, not only at the finished artifact: 89% of users turned to it for finding information and research, and the most useful tasks — summarizing, step-by-step instruction, brainstorming — overlap directly with coursework activities.
• Instructors. Teach AI judgment explicitly instead of leaving students to improvise it: 95% learned GenAI use through trial and error, 92.5% through internet searches and 88.7% by using GenAI itself, while university workshops and online resources were the least popular and least helpful.
• Administrators. Write policy for legitimate use rather than blanket prohibition: 68% of respondents never used GenAI when it was not permitted and another 21% only sometimes, against 5.2% who used it most of the time or always.
• Administrators. Respond to refusal as a reasoned position rather than a gap to close: 32% of respondents did not use GenAI for study, and non-users cited wanting to do the work themselves (86%), inaccuracy (81%) and broader ethical and environmental concerns.
• Researchers. Test access assumptions before treating GenAI as a universal aid: disabled students were regular users at 35% versus 52% of non-disabled peers, and accessibility was rarely cited as a motivation, so report use by subgroup rather than as one headline figure.

Limitations

• The convenience sample at 7 institutions is not representative of UK higher education as a whole: institutional response rates varied sharply, and non-response bias may skew results toward students with strong views.
• The UK survey was administered at the start of the academic year, when first-year and master's students may be less familiar with institutional approaches, whereas the Australian survey ran near the end of the year.
• Self-report surveys capture behavioral intentions rather than actual actions, with social desirability and contextual pressures likely to under-report academic misconduct.

Citation

Gow, S., Illingworth, S., Fabian, K., & Goddard, C. (2026). "It is a temptation to get it to do the work…" Student experiences of navigating the generative AI landscape in UK higher education: A cross-institutional survey with international comparison. EdArXiv preprint.