Authentic assessment is central to educational certification: a qualification is credible only when assessment gives defensible evidence of what a student knows and can do. Yet cheating is as old as assessment itself. In imperial China, candidates in civil service examinations used tiny “fly head” script notes; Princeton University’s Cotsen Children’s Library has documented a surviving example. The tools may have gotten more advanced; the incentive or desire to gain an unfair advantage has not.
In my own teaching at the University of Amsterdam, I have felt compelled to replace take home assignments in my courses with proctored exams in an effort to enable authentic assessment. This decision was not at all premised on me opposing students’ leveraging AI, but rather to counter some students’ tendency to outsource their learning process to AI. I can imagine it must be a struggle for at least some students to invest effort in learning, when assessment outcomes and associated products can be generated at the click of a button.
The university has also implemented changes to thesis supervision in which master’s students are now being asked to orally present the progress on their theses at several points in the research process. This means that instead of evaluating a polished final product, thesis supervisors are now much more aware of individual students’ deliberations as they develop their work, and their ability to account for the decisions they made in the process. But neither a proctored exam nor a series of progress presentations guarantee valid assessment. Furthermore, it is oftentimes challenging to the point of being impossible to prove beyond a reasonable doubt that a student has illicitly used AI, when they themselves maintain that they did not. In this case I think educators should aim to err on the side of caution, as sanctioning a student for an act they did not commit may have graver consequences than not catching a student who did. Furthermore, proctoring and AI policing more generally can narrow what we measure and require substantial time and resources. The take home assignment I ended up dropping from my course, for instance, was a group assignment in which students used to have opportunities to hone their teamwork and collaboration skills. Indeed, I very much experience these decisions as imperfect tradeoffs, and will continue to invest in developing novel assessment methods that encourage and support students’ acquisition of knowledge, skills, and abilities.
Doing so requires at minimum an awareness of the full range of methods and tools out there that may be leveraged to compromise authentic assessment, and it is here that many educators, myself included, may be falling behind the rapid technological development. Contemporary risks combine generative AI, answer sharing platforms, contract cheating, remote collaboration, consumer devices, and services that can produce or revise work quickly. Recent risk assessment research asks which tasks generative AI can complete and what evidence can still establish individual learning (Newton, 2025). A 2024 Inside Higher Ed report reflected growing concern among students and faculty that AI makes cheating easier. Still, AI use is not automatically cheating: the relevant questions are what the assignment permits, what students disclose, and whether the submitted work demonstrates the intended learning outcomes.
Overview of risk of cheating by assessment format
| Assessment format | Broad means and risks | Indicative cost / effort |
| In person, proctored, closed book questions | Unauthorised notes or devices, collusion, or impersonation. Examples of hardware include a smartphone, smartwatch, wireless earbud, smart glasses, a second phone, USB storage device or programmable calculator. A helper may relay answers or impersonate. | Often little added cost if a device is already owned; buying dedicated electronics can cost tens to hundreds of euros. Coordination and detection risks are higher than in unsupervised work. |
| Proctored, open book or open resource questions | Permitted resources reduce the value of hidden notes, but unauthorised AI, messaging, a second device or another person’s input can still replace the student’s reasoning. | Usually low direct cost if a device or free software is already available. The main risk is unauthorised assistance, not just the presence of resources. |
| Unproctored online tests, closed or open resource | Search, generative AI, shared answer banks, group messaging, a helper completing items, or identity substitution. A second phone or laptop can facilitate collaboration; software may include chatbots or messaging tools. | Often free access or variable monthly subscription costs for consumer software. Human help may be informal or paid; pricing varies widely. |
| Take home essays, problem sets and projects | AI can draft, explain, translate or code; copied work can be adapted; peers or paid third parties can complete some or all of a task. The final product may not reveal how much assistance was used. | Free tiers to monthly subscriptions costing tens of euros; commissioned work can range from informal assistance to hundreds of euros or more. Highly variable. |
| In person or synchronous video presentations | Slides or analysis may be substantially prepared by someone else; a student may receive prompts or use a proxy. Remote presentations add identity verification risks. Hardware/software can include a second device and messaging; helpers can act as coaches or ghostwriters. | Often negligible if devices are already owned; coordination is the main investment. Live questions make it harder, but not impossible, to sustain a false claim of ownership. |
| Asynchronous video presentations | Scripts can be generated by AI or written by someone else; editing can hide weak command of the material. Synthetic video or voice can potentially impersonate a student or misrepresent who made a submission. | Free editing or AI tools to paid subscriptions; outside production assistance varies widely. A polished recording is easier to prepare than to defend spontaneously. |
Deepfakes deserve attention, but the evidence should be described carefully. Synthetic face or voice could be used to impersonate a candidate in a remote oral assessment, or to make an asynchronous recording appear to come from someone who did not produce it. This is a plausible and evolving identity risk; it should not be presented as a proven, widespread method of university exam cheating without stronger evidence. The same caution applies to other sophisticated threats: the fact that something is technically possible does not prove that it is being used in practice.
Likewise, assessment types should not be ranked as simply “safe” or “unsafe.” A closed book exam limits outside access but can overemphasize recall. An open book task may assess application while leaving room for unauthorized help. A take home project may be authentic in its subject matter, yet weak at ascertaining authenticity on the part of who produced it. Oral assessment can reveal understanding, but may privilege speed, confidence or spoken fluency and disadvantage some students. In sum, “authentic” is not synonymous with “immune to cheating.”
What educators can change
From a constructive alignment angle, it may be best to start with the learning outcome, before deciding on the assessment form and appropriate safeguards. If rapid recall under constraint is essential, supervised testing can be appropriate. If application matters, use unfamiliar cases, data or scenarios and ask students to justify their reasoning. Open book questions can be more meaningful when they require interpretation or comparison rather than reproduction. Question banks and variants may help at scale, but must preserve comparable difficulty and fairness.
Gather evidence of process without turning learning into surveillance. Staged proposals, annotated sources, drafts, data logs, code histories, reflections and progress conversations can help show how a project developed. My thesis progress presentations let me ask why a design changed, what evidence supports a claim and what comes next. Such checkpoints strengthen the evidence when considered together; none is conclusive on its own. Requiring exhaustive keystroke histories or constant monitoring may be intrusive and can end up assessing compliance as opposed to learning.
Be transparent about the limits of proving AI use. For many take home assignments, it may be impossible to establish from the submitted text alone whether a student used generative AI, how extensively they used it, or whether that use crossed a line. Human and AI contributions can be blended through brainstorming, translation, rewriting and editing; the same polished prose can have several plausible origins. AI detectors produce uncertain classifications, not proof of who authored a particular passage or what the student intended. Liang and colleagues (2023) found that detectors disproportionately misclassified non native English writers in their study. To date, a detector flag should not stand alone as evidence for an accusation. A fair process needs a clear rule, an opportunity for the student to explain, and corroborating evidence where available, while recognizing that process records and oral follow ups also have limits.
Use oral interaction selectively and fairly. A brief follow up question, asked of everyone or a transparent sample, can help verify individual understanding after an essay or presentation. Synchronous video can serve remote cohorts, but adds identity and prompting concerns; in person questioning has its own access and scheduling constraints. Asynchronous video is scalable, but where independent mastery is central it may need another evidence source. Rubrics, advance notice, accommodations and sampled follow ups can reduce workload and arbitrariness.
Explicate permitted and prohibited AI use expectations. Be clear about whether brainstorming, grammar support, translation, coding assistance or generative drafting is permitted, and what must be acknowledged. Some courses should assess responsible AI use; others need evidence of unaided performance. Avoid blanket prohibitions that are impossible to enforce consistently, as well as vague permissions that leave students unsure as to what counts as misconduct. UNESCO’s guidance recommends a human centred approach to AI in education rather than treating adoption or prohibition as a complete policy.
Treat surveillance as a limited tool, not a solution. Remote proctoring may deter some behavior, but it cannot guarantee identity or prevent every unauthorized resource. It can also raise privacy, accessibility, technical, and equity concerns. More proctoring is not automatically more valid assessment.
Finally, cost and scale matter. Proctored exams require rooms, staff, scheduling and accessible alternatives; individual oral defenses offer rich evidence but can be expensive for large cohorts. Automated tests scale more readily but need good question design and moderation. A practical portfolio might combine a supervised assessment of individual competence, take home work with process checkpoints, and selective oral verification. The goal should not be total control at any cost; it is enough to have credible evidence for the decision being made.
Conclusion
Cheating is a persistent pressure on certification, not a new invention of AI. What has changed is the efficiency and effectiveness with which software, devices, and services can produce work or help another person appear to have produced it. Educators cannot eliminate every opportunity without making assessment intrusive, expensive or less educationally valuable. We can make more defensible choices: align the format with learning outcomes, gather evidence at multiple points, state assistance rules clearly, and avoid treating suspicion or a detector score as proof. The aim is not to catch every possible cheat. It is to make assessment a fair, proportionate and credible account of what each student can do.
Stefan Mol
Note on AI assisted Writing:
The conceptualization and ideas contained within this article are the author’s. Yet, AI assistants helped write the first draft and formulate broad ideas into precise words. Ideas and concepts were evaluated, modified, refined, and checked.
References
Newton, P. M. (2025). How vulnerable are UK universities to cheating with new GenAI tools? A pragmatic risk assessment. Assessment & Evaluation in Higher Education, 50(8), 1332 to 1343.
Inside Higher Ed. (2024, July 29). “Students and professors expect more cheating thanks to AI.” https://www.insidehighered.com/news/tech-innovation/artificial-intelligence/2024/07/29/students-and-professors-expect-more
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non native English writers. Patterns, 4(7), 100779.
Chen, M. (2019). Cheating in Examinations for Cheapskates? A Centuries Old Tip from the Chinese Collection of the Cotsen Children’s Library.
UNESCO. (2023). Guidance for generative AI in education and research. https://unesdoc.unesco.org/ark:/48223/pf0000386693
