AI-based digital cheating at university, and the case for new ethical pedagogies

Synthesis: Leaton Gray, Edsall and Parapadakis argue that generative AI has not created a crisis of Academic Integrity so much as exposed one: massified, standardized and depersonalized assessment regimes already measured little more than the ability to locate, compile and paraphrase information, which is exactly what large language models now do cheaply and convincingly. Working from Ajzen's Theory of Planned Behavior, Bandura's self-efficacy theory and situational crime prevention, they ask how far assessment practice creates the conditions for AI-facilitated misconduct and how ethical pedagogy might mitigate it. Their answer is grounded in the published case of Baird and Clare (2017), an Australian business capstone where 25 situational prevention techniques reportedly cut misconduct cases from 183 to 27 within a year. Detection and online proctoring, they conclude, are unreliable and anti-educational, since governance built on suspicion deepens estrangement between students and institutions; the sustainable path is assessment redesigned around authenticity, personalization and iterative engagement.

Core Argument

1. AI lowers the cost of misconduct that was already endemic. AI-generated text has passed as human-authored in up to 80% of cases (Cotton et al., 2023; Eke, 2023; Elkhatat et al., 2023; Liu et al., 2024), and LLMs simply join the 'e-cheating' methods catalogued by Dawson (2020) before generative AI arrived. Cheating studies reach back to Campbell (1935) and Bowers (1964), and Murdock and Anderman (2006) put the share of students who had cheated at secondary school and university as high as 80–90%. AI amplifies an existing condition, which is why banning tools cannot reach its source.
2. Cheating is planned behavior, driven by motivation and perceived control. Combining Ajzen's (1991, 2005) Theory of Planned Behavior with Bandura's (1977) theory, the authors trace an external locus of control linked to dishonesty (Leming, 1980) and an internal locus to honesty (Rinn et al., 2014): when students feel little Learner Agency, the perceived barriers to cheating fall. Murdock and Anderman's (2006) model frames the decision as three questions: "What is my purpose?", "Can I do this?", "What are the costs?". The counter-intuitive evidence is Krou et al.'s (2021) meta-analysis, where Self-Efficacy correlates negatively with cheating while actual ability does not inversely correlate with it at all: a capable, confident student who reads an assessment as unfair may cheat to regain control.
3. Detection and proctoring are a losing position. Transformer models paraphrase rather than copy, so their output matches nothing in a plagiarism database, and the evidence cited puts machine detection at about 80% against 78.4% for human reviewers (Wahle et al., 2022), too narrow a margin to carry a misconduct finding. Digital proctoring, delivered through tools such as Turnitin and Safe Exam Browser, relies on static surveillance and basic liveness checks that can be bypassed, assumes students are potential cheaters rather than active Learners (Lee and Fanguy, 2022), and is invasive enough to become a target for subversion (Simko et al., 2024).
4. An assessment that AI can answer convincingly is an intellectually trivial assessment. If a model can source information, assemble a coherent narrative and paraphrase it without understanding the subject, and thereby deceive an assessor, what is measured is keyword identification and compilation, not knowledge or ability. The failure belongs to the assessment, not the student, and predates AI: the technology has merely exposed the superficiality of standardized, anonymously marked tasks designed for administrative convenience rather than pedagogical fidelity (Blackie, 2024; Kramm and McKenna, 2023).
5. Institutional procedure makes enforcement unlikely even when cheating is suspected. Quality assurance protocols demand forensic evidence at a level that deters staff from pursuing investigations (Brigham and Ziebart, 2020; Keith-Spiegel et al., 1998), yet unaddressed cheating tends to proliferate (Packalen and Rowbotham, 2022). Competition among contract cheating providers, including platforms that use AI to automate client communication, has cut the cost of commissioned work inside the 'gig academy' (Gaumann and Veale, 2024; Kezar et al., 2019; Sweeney, 2023). The result is a war of attrition (Keir and Ives, 2022) that lecturers cannot win by spending more time on detection as cohorts grow.
6. Cheating is not generally seen as inherently immoral, and AI leaves its social costs untouched. The authors find it "surprisingly" clear that cheating does not evoke strong moral outrage among students (Ashworth et al., 1997; Marsden et al., 2005), a view shared by academics (Godecharle et al., 2018; Martinson et al., 2005) and society at large (Henle et al., 2019; Wood et al., 2007). AI reduces the time, effort and money cheating costs but barely touches its social costs, such as humiliation and ostracism, and that asymmetry is the opening for pedagogy: where students identify with an aspirational ideal future self (Boyatzis and Akrivou, 2006) as members of an intellectual community that rejects cheating, the behavior can be discouraged before it is planned.

Theory, method and what the article is

This is a conceptual article rather than an empirical study. It poses two research questions, on how far current assessment practice breeds AI-facilitated misconduct and how validity-based ethical pedagogy might mitigate it, builds its argument from the Theory of Planned Behavior, self-efficacy theory and situational crime prevention (Clarke, 2017), and grounds it in a case drawn from the published literature rather than data the authors collected. They also say which theories they set aside: General Strain Theory (Agnew and Brezina, 2019) and Neutralization Theory (Sykes and Matza, 1957) address strain most students do not face, while Routine Activity Theory (Cohen and Felson, 1998) and Social Learning Theory capture less of the individual decision-making involved. A footnote concedes that external factors such as variable teaching quality and poor learning environments were left untheorised, and the authors add that no cross-culturally stable definition of cheating exists, since some systems emphasize memorization and reproduction while others require critical interpretation and originality (Wanyama Wanyonyi, 2024).

The Baird and Clare case: prevention instead of punishment

The article's concrete evidence comes from Baird and Clare's (2017) crime prevention case study of an Australian university business capstone, in which 25 situational crime prevention techniques adapted from Clarke (2017) were applied to Assessment design so as to raise the effort and risk of cheating, reduce its rewards, reduce provocations and remove excuses. Reported misconduct cases fell from 183 to 27 within a year. The case does not concern AI or LLM use directly, but the authors treat it as a framework for the social and pedagogical conditions that do apply.

The mechanisms are what the article's proposals generalize from. A computer-based Simulation tracked student interactions; a red-flag system detected discrepancies and work that looked too expert; random reassignment of team members, a "board shake-up", raised both effort and risk. Students received easily digestible information on misconduct standards, which the case authors describe as priming a student's conscience, and course materials that students had published online, in some cases for payment, were met with takedown notices under copyright law. Weekly in-class invigilated tests raised the effort required to cheat while restoring personal oversight and lecturer-student interaction. The limits are recorded too: granular tracking suited a structured simulation but may not transfer to essays or project work, and bespoke single-use materials conflict with current workload models.

Rehumanizing assessment: what the authors propose

The prescriptions follow from the motivational analysis rather than from detection. Universities should raise the perceived purpose of assessment, build self-efficacy, and increase the perceived social cost of cheating, and the authors warn that addressing only one or two of these will fail because the model needs all three. Dawson et al.'s (2024) validity framework is presented as the structured route to aligning assessment with intrinsic motivation, and the article offers five discipline-specific redesigns:

• STEM: timed problem-solving exams become open-book problem solving with a written reflection on the student's own process.
• Humanities: summative essays become collaborative research projects using primary and secondary sources, such as a group digital exhibit built on archival research.
• Business: case study analysis becomes real-world simulation with peer-reviewed presentations, such as a market strategy pitched to peers and instructors.
• Social Sciences: theoretical essays become community-based action research with local partner organizations.
• Creative Arts: portfolio submissions become iterative peer-reviewed design processes with Feedback from peers and industry professionals.

The common commitments are authentic real-world tasks, personalization to the student's interests, and process-oriented work that carries feedback through the middle of a task rather than only at its end. Formative assessment is the corrective to the depersonalized lecture-and-exam pattern massification produced, and the authors draw on pandemic experience of remote teaching, where low-stakes iterative work and weekly collaborative assignments increased intrinsic motivation and ownership (Surahman and Wang, 2022). Viva-style examinations and peer-mediated assessment are argued to reduce reliance on AI tools, and digital divides in devices, connectivity and caring responsibilities must be addressed, since equitable access is a precondition of integrity.

What this means for practice

• Educators. Treat an assignment a model can answer convincingly as a redesign prompt rather than a misconduct problem: the article's discipline-specific replacements include turning timed problem-solving exams into open-book problem solving with a written reflection on the student's own process, and summative essays into collaborative research projects built on primary sources.
• Educators. Move all three motivational levers at once — the perceived purpose of Assessment, students' Self-Efficacy, and the perceived social cost of cheating — because the authors state plainly that addressing only one or two of them will fail.
• Administrators. Stop funding detection and proctoring as the integrity strategy: transformer output paraphrases rather than copies so it matches nothing in a plagiarism database, cited machine detection runs about 80% against 78.4% for human reviewers, and governance built on suspicion damages Trust in ways the authors say the academy will not easily repair.
• Researchers. Test the validity-based redesigns and the motivation pathway this article theorizes but does not measure — validity-based interventions across disciplines and cultures, and the self-efficacy account of cheating the article builds from Krou et al.'s meta-analysis.

Limitations

• This is a conceptual article, not an empirical study: no data were collected, the argument is built from Ajzen's Theory of Planned Behavior, Bandura's self-efficacy theory and situational crime prevention, and the authors themselves call for empirical work testing validity-based interventions across disciplines and cultures.
• Its only quantitative evidence is one published case study — Baird and Clare's (2017) Australian business capstone, where 25 situational prevention techniques were followed by a fall in misconduct cases from 183 to 27 within a year. That result is reported by that study's authors rather than replicated, is not about AI use, and its granular simulation tracking may not transfer to essays or project work.
• The proposed alternative to detection is resource-intensive in exactly the direction universities are economizing: bespoke single-use materials, weekly in-class invigilated tests and personalized feedback all consume staff time, and the authors concede this conflict with current workload models.
• The theoretical frame is partial by the authors' own account — a footnote concedes that external factors such as variable teaching quality and poor learning environments were left untheorized — and they note that no cross-culturally stable definition of cheating exists, since some systems emphasize memorization and reproduction while others require critical interpretation.

Citation

Leaton Gray, S., Edsall, D., & Parapadakis, D. (2025). AI-based digital cheating at university, and the case for new ethical pedagogies. Journal of Academic Ethics, 23, 2069–2086.