Corroboration, documentation, and what a student's work can still be made to show in the AI era
Every assessment asks two questions of a piece of work: who wrote this, and what was learned? Generative AI has made both harder to answer from the finished text alone. A polished submission no longer reliably shows what a student can do, and no single check can tell us who did the work.
In May 2025 we published our first white paper, Academic Integrity in the Age of AI: A Practical Guide. Since then, the debate has moved on, with new data, new court rulings and regulators who have started to look at how work develops, not only at what is handed in. Our 2026 edition, Academic Integrity Beyond Detection, is written for university leaders, examination officers, educators in higher education and schools, and policymakers.
Table of Contents
ToggleWhat has changed since 2025
Use is near universal; misuse is narrower, but real. In the 2026 HEPI survey of UK undergraduates, 95% use AI in some form, and 12% say they have put AI-generated text directly into assessed work, up from 3% in 2024. A study of more than 95,000 students at 20 US public research universities estimated that 9% had used generative AI to cheat, rising to 26% among daily users.
The deeper issue is validity. In a randomised study, AI support improved students' essays without improving what they learned. In a school field experiment, unrestricted GPT-4 raised practice scores but lowered later unaided exam results by 17%; a version that gave hints instead of answers largely avoided the harm. At Harvard, a carefully designed AI tutor produced more than twice the learning gains of active learning in class. How AI is used decides whether it supports learning or replaces it, and the finished text rarely shows which.
None of this is an argument against AI. Used well, it is a real benefit to education, and it is here to stay. The aim is work written with AI, not by AI. Whether help comes from AI, a published source or another person, the questions stay the same: was it allowed, and was it declared?
Why no single check settles authorship
Each familiar check answers a narrower question than the one assessment needs answered: what did this student do? Classic similarity checking struggles with paraphrase and translation. Declarations do not verify themselves. A watermark can show that a model was involved, but its absence proves nothing.
AI-text detectors raise a sharper problem. On long, unedited human texts, the best tools now rarely raise false alarms. But in a 2026 study from the University of Notre Dame, abstracts lightly edited with AI were flagged in 38% to 80% of cases, while text run through a commercial rewriting tool was missed at least 96% of the time. The authors conclude: "Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion."
At the heart of the paper: the corroboration standard
If no single check settles authorship, procedures should not pretend otherwise. The paper proposes a simple rule: no finding rests on a single tool's output. A finding needs converging evidence from at least two of four categories, a fair opportunity for the student to explain their work, and a human decision.
- Textual: similarity and semantic source matching, AI classification, citation checks, stylometric analysis.
- Process: dated drafts, version history, research trail, declared AI use with chat logs.
- Conversational: the student's explanation of their choices in a short oral check.
- Contextual: course participation, supervision notes, the conditions under which the work was set.
This reflects existing law more than it adds to it. German case law, the Office of the Independent Adjudicator for England and Wales and data protection law point the same way: the institution carries the burden of proof, and an automated score must not decide alone. The paper walks through recent rulings from Germany and New York.
The same discipline applies to judging learning. A long research trail may reflect confusion, and a clean draft may reflect transcription. Process data must never be read as a competence score. A finding is not a verdict.
Why documenting the process deserves a central place
Many institutions want to keep extended, unsupervised work such as theses, and for good reason: writing is a way of learning, and months of planning, managing sources and revising an argument train something different from a single exam. Monika Oertner's model of academic core competencies shows what is at stake. The paper's advice: keep the project, let AI support it where the task allows, and count the path as evidence, not only the product.
Documenting that path produces evidence during the work, not months later under suspicion. It gives honest students a record of their effort, makes AI-permitted tasks assessable and shows a stalled project while help is still useful. One record serves both questions: whose work this is, and how the learning developed.
The paper is just as clear about the limits. A record cannot prove authorship on its own (faking one is possible, though laborious), it cannot see work done elsewhere, and its outcomes have not yet been independently evaluated. That is a reason to pilot and evaluate, not to wait. Our article on process documentation as an assessment format shows what that can look like in a course.
What you will find inside
- Ten approaches compared, from AI rules and oral exams to source matching and process documentation, with their costs and evidence.
- Seven principles for responsible implementation, including readiness for the EU AI Act's high-risk rules for education from December 2027.
- A seven-step roadmap and a checklist for pilots.
- 43 references across Europe, the UK, the US and Australia, from peer-reviewed studies to court rulings and regulator guidance.
Most first steps need no new software. And in keeping with its own argument, the paper ends with a note on how we used AI in writing it.
About Mentafy: how our tools fit this approach
This section is about our own products. The argument of the white paper stands without them.
Mentafy Check works on finished submissions: similarity checking, Semantic Source Search (S³) for paraphrased and translated sources, a Reference Verifier, optional checks of submissions against each other, and an AI-text classifier that we treat as a triage signal, never as evidence of authorship. Mentafy Trail documents work as it develops, from document snapshots rather than keystroke logging, with a Research Recorder and a record students can see and annotate. Whether you check finished work, document the process, or both, every signal feeds the same evidence base: who wrote this, and what was learned.
As the paper puts it, independent, published evaluation will do more for the field than any vendor claim, including ours. We welcome any institution that wants to run and publish one.
Selected sources
- Stephenson, R., & Armstrong, C. (2026): Student Generative AI Survey 2026, HEPI Report 199, Higher Education Policy Institute and Kortext.
- Chirikov, I., Smirnov, I., & Kizilcec, R. F. (2026): Generative AI use and misuse call for assessment reform in higher education, Science 392(6800), 818-820.
- Fan, Y., et al. (2025): Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance, British Journal of Educational Technology.
- Bastani, H., et al. (2025): Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS 122(26).
- Kestin, G., et al. (2025): AI tutoring outperforms in-class active learning, Scientific Reports 15, 17458.
- Karr, J. A., Jr., Khvatskii, G., Hua, T., & Chawla, N. V. (2026): Why AI detection fails for academic integrity, ACM AI Summit 2026 (preprint).
- Office of the Independent Adjudicator for Higher Education (2025): Casework note: Complaints relating to AI and academic misconduct, July 2025.
- VG Kassel, judgments of 25 February 2026, 7 K 2134/24.KS and 7 K 2515/25.KS; Matter of Newby v Adelphi University, 2026 NY Slip Op 26021 (Sup. Ct. Nassau County).
- Full references are in the white paper.






No comment yet, add your voice below!