Academic Integrity: Why AI Detectors Are Not Enough

Faced with generative AI, many institutions' first reaction was defensive: detect machine-produced text. The intent is understandable, but the research is damning — today's detectors do not deliver on their promises, and relying on them can do more harm than good.

Unreliable detectors

A landmark study tested fourteen detection tools: they are judged 'neither accurate nor reliable', tend to classify AI text as human, and are easily defeated by paraphrasing or obfuscation ^[1]. Circumvention is trivial; the false sense of security, dangerous.

The consequences are not theoretical. Investigations have documented international students falsely accused of cheating on the strength of such tools ^[3]. The market's leading vendor itself acknowledged higher false-positive rates than its lab figures, especially for documents with little suspect text ^[4].

Detection alone is a dead end

A systematic review of 41 studies concludes that generative AI boosts engagement and efficiency but poses real dishonesty risks — and that detection-only approaches are insufficient ^[5]. Teachers themselves rate anti-AI policies as less effective than traditional rules and prefer educative over punitive responses ^[6].

What works: rethinking assessment

• Design AI-resistant assessment: oral, process, drafts, local grounding.
• Make AI use explicit and documented rather than clandestine.
• Teach AI literacy: when, how and why to use it honestly.
• Favour remediation and dialogue over automatic sanction.

Frameworks are emerging along these lines, shifting integrity from policing toward education, coaching and remediation ^[7]. UNESCO, for its part, argues for a human-centred approach and clear national policies ^[8]. Integrity cannot be decreed by a probability score: it is built through task design and trust.

You do not restore trust by accusing the innocent; you build it by making cheating pointless and learning desirable.

Learnya synthesis

Toward a culture of integrity, not an arms race

Betting on detection means entering an arms race lost in advance: every improvement in detectors is immediately circumvented by a paraphrase or a new model, while the false positives strike the innocent ^[1]. This asymmetry should be enough to disqualify detection as a pillar of integrity policy. At best it can serve as one indicator among others in a dialogue, never as automatic proof.

The durable answer is cultural and pedagogical. A culture of integrity is built by making expectations explicit (what is acceptable AI use in this course?), by valuing process as much as product, and by handling breaches through dialogue and remediation rather than sanction alone. Teachers themselves favour these educative approaches ^[6][7], more effective in the long run than fear of the algorithmic police.

Rethinking assessment plays a central role. Tasks grounded in a local context, oral defences, and work that requires showing one's reasoning make cheating both harder and less interesting than learning itself. Far from a constraint, the arrival of AI offers a chance to rebalance assessment toward what really matters: the ability to think, to justify and to create.

Integrity as learning, not suspicion

The most fruitful shift is to see integrity not as a policing problem, but as an object of learning. Understanding why to cite sources, what an honest contribution is, how to declare the help received from an AI: these are ethical and professional skills that prepare students well beyond school. Treating them this way turns a threat into an educational opportunity.

Concretely, we can anticipate the spread of explicit use charters, logbooks documenting AI use, and assessments that value the journey as much as the result. This framework better protects learners — including the most vulnerable, unfairly exposed to detectors' false positives — while restoring a trust that generalised suspicion can only destroy.

A practical example makes the shift tangible. Replacing a take-home essay with a short oral defence of that essay does not require any detector: it simply asks the student to demonstrate ownership of their own reasoning. Combined with drafts, process logs and locally grounded prompts, such designs make dishonesty both harder and less rewarding than genuine work. The most durable safeguard against misuse is an assessment worth doing honestly — one that measures thinking the student cannot outsource without revealing it.