The Impact of Generative AI on Academic Integrity of Authentic Assessments Within a Higher Education Context

Synthesis: Kofinas, Tsay, and Pike (2025) ran a series of experiments across two UK universities in which experienced academic markers judged human-authored, GenAI-modified, and GenAI-generated assessments. They found that markers generally could not distinguish assessments with GenAI input from those without, that the presence of GenAI nonetheless affected how markers approached marking, and that the level of authenticity in an assessment had no impact on the ability to safeguard against or detect GenAI use. The authors conclude that current assessment approaches are susceptible to GenAI manipulation and that authentic assessments alone cannot protect academic integrity; the focus must shift to assessment design favoring process-based, performative, and synchronous interpersonal forms.

Key Findings

1. Markers cannot reliably detect GenAI. Markers generally could not distinguish assessments that had GenAI input from those that did not, even when the assessment was authentic.
2. GenAI alters the marking process. The mere possibility that GenAI was used (and suspicion of tampering) changed markers' behavior and judgment, producing both false positives and false negatives—so students' learning is not assessed correctly.
3. Authenticity offers no safeguarding benefit. The level of authenticity in an assessment had very limited impact on the ease of detecting GenAI usage; AI could generate authentic assessments that passed faculty scrutiny.
4. Consequentialist reasoning drives misuse. Students weigh the benefits (efficiency, ease) against risks of getting caught; when benefits outweigh risks, they may use GenAI in ways that compromise integrity.
5. GenAI assessment is cheap and 24/7. Unlike essay mills or contract cheating, GenAI can produce high-quality, seemingly original work for free or very cheaply, undermining authorship detection and circumventing the effort required to learn.
6. Output-based written assessments are vulnerable. Common written formats—reports, essays, and take-home exams—are susceptible, whereas performative, synchronous, and socially experiential assessment formats are more resistant.

What this means for practice

• Assessment designers. Stop treating authenticity as a safeguard: the level of authenticity in an assessment had very limited impact on detecting GenAI use, and GenAI-generated work passed faculty scrutiny — shift the effort to process-based, performative, and synchronous interpersonal designs instead.
• Instructors. Account for the awareness effect when marking: knowing GenAI might be present produced six false positives on unaltered student work and seven missed AI samples, and markers assigned high-scoring samples grades up to 10 percentage points lower.
• Institutions. Revisit marking, moderation, and credentialing for written assessments now that markers generally cannot distinguish GenAI-modified from human-authored submissions — reports, essays, and take-home exams are the most exposed formats.
• Assessment designers. Build authentic, presence-required formats that require students to defend work in the moment, such as synchronous oral and dialogic assessment, and treat experiential and performative tasks as the alternatives that resist GenAI manipulation, complementing the design-for-presence agenda in agentic AI debates.
• Instructors. Plan for false negatives as well as false positives: markers who are told GenAI may be present change how they grade, so assessment validity — not detection — is the design problem.

Limitations

• The experiment drew on only eight markers (four academic pairs) at two UK universities, evaluating 21 randomly coded writing samples across three undergraduate levels; the authors describe it as exploratory with a small sample.
• Detection performance varied sharply between pairs — Pair 2 reached 33% accuracy against Pair 1's 85.7% — and Pair 1's participants included the module leader for the levels the samples came from, so accuracy may reflect familiarity rather than general marker capability.
• Markers were told in advance that some samples had GenAI input, a condition unlike normal marking that plausibly induced suspicion effects, and the design used pairwise consensus reconciliation with no third marker or moderator.
• The six assessments sampled were all judged low in business proximity (none conducted in an actual workplace), and authenticity was mapped to two frameworks post hoc rather than varied experimentally.

Citation

Kofinas, A. K., Tsay, C. H.-H., & Pike, D. (2025). The impact of generative AI on academic integrity of authentic assessments within a higher education context. British Journal of Educational Technology, 56(6).