Beyond Detection: How Students Use—and Hide—AI in Online Assessments and What Authentic Tasks Can Do About It

Synthesis: Students navigate institutional AI ambiguity with caution and pragmatism — disclosure is rare, driven more by fear of penalties than by dishonesty. Kirsanov, Kushwah, and Selvaretnam (2025/2026), a small case study of undergraduate economics students at the University of Glasgow (31/174 respondents), find that only about one-third report using AI in online assessments, and disclosure is rarer still. Non-disclosure appears to be rational caution in the face of ambiguous policies and perceived academic risk, not simple integrity violation. Students support guidance and structured AI Regulation in Education, favor citation rules, and widely see real-world, data-based tasks as reducing AI misuse — pointing toward authentic assessment and clear expectations over detection and deterrence.

Key Findings

• Reported AI use is moderate, and disclosure is rare. Only ~34% of respondents acknowledged AI use; most uses were supportive (proofreading, rewording, brainstorming), and no student reported submitting a full AI-generated answer. Many opted out of reporting AI use early in the survey — interpreted as strategic self-presentation rather than proof of non-use.
• Fear of penalties is the dominant barrier to disclosure. About three-quarters of respondents worried about being penalized for AI use, even when use was moderate or constructive ("penalty anxiety"). This is consistent with broader UK/international survey evidence.
• Ethical anxiety predicts lower disclosure. Logistic regression found students with greater ethical anxiety (fear of penalties, unfair access, undermining learning) were less likely to disclose AI use. Notably, UK nationals disclosed significantly more than international students.
• Students want guidance, not bans. Most agreed AI is acceptable when used supportively, and many welcomed clearer citation rules and institutional guidance. Views on a blanket ban were divided.
• Authentic, data-rich tasks are the promising lever. Students widely supported real-world, data-based tasks as reducing over-reliance on AI — by design rather than deterrence.

What this means for practice

• Assessment designers. Design real-world, data-based tasks: students widely endorsed them as a way to reduce over-reliance on AI, making task design rather than deterrence the lever.
• Educators. State explicitly that honest disclosure will not be penalized — about three-quarters of respondents feared penalties for AI use even when that use was moderate or constructive.
• Administrators. Publish clear citation and use guidance so students have a framework and the confidence to disclose voluntarily instead of navigating ambiguous policy alone.
• Educators. Address the access and understanding gap behind disclosure: UK nationals disclosed significantly more than international students, and only about one-third of respondents reported any AI use in online assessments.
• Administrators. Shift resources from detection and deterrence toward guidance and structured regulation, which these students supported; non-disclosure here read as rational caution under ambiguous policy rather than dishonesty.

Limitations

• Small single-program sample: 31 respondents, about 18% of the cohort (31/174), at one UK economics program, with potential self-selection bias; the authors state that the findings are exploratory and not generalizable beyond this setting.
• Female and international students were somewhat over-represented among respondents relative to the cohort as a whole, which may also have shaped the patterns observed.
• All measures are self-reported, so social desirability bias is a live risk — students may have underreported AI use or overstated ethical caution despite the survey's anonymity.
• Perceptions were captured at a single point in time, before the release of more advanced AI tools and detection systems, and the small sample size limits the strength of any statistical claims from the exploratory regression.

Citation

Kirsanov, O., Kushwah, L., & Selvaretnam, G. (2026). Beyond Detection: How Students Use—and Hide—AI in Online Assessments and What Authentic Tasks Can Do About It. Journal of Academic Ethics, 24, 14.