Currently reading Bassett et al. (2026) on “Heads we win, tails you lose: AI detectors in education”

AI detection products cannot produce output to reliably destinguish between an AI generated text and one written by a human, and they reflect biases (for example more often flagging non-native speaker texts — see also my recent post about the “GenAI writes like me” post). We’ve known that for a while now, so why summarize another article on AI detection?

Because Bassett et al. (2026) is a really useful article!

They summarize why AI detection products cannot work in the first place:

• Unverifiable probabilistic estimates. Usually, one should be able to distinguish detectior outputs between “hit (true positive), miss (false negative), false alarm (false positive), or correct rejection (true negative)“, and from that understand the detector’s sensitivity. But for AI detection, there is no ground truth to verify the model, since the only texts that are definitely not written using AI are old and it is not a valid assumption that texts written today still follow the same linguistic and stylistic pattern (since students might have learned from AI generated texts, and language also just evolves over time).
• Mutually exclusive linguistic markers. “There is no principled reason to believe that a human cannot produce writing that contains linguistic features commonly found in AI-generated text” (especially since AI is trained on human writing, and as humans are exposed to AI texts and might adapt to them)

They also describe common methods that teachers use to validate AI detector outputs and why they don’t work, either:

• Linguistic markers: it sounds too perfect, or uses the type of lists or words that people associate with AI (“thus”): “staff reinforce their assumptions using reasoning that is entirely dependent on the AI detector’s potentially flawed output“
• Multiple AI detectors: which only “amplifies the shared flaws of these tools, creating a misleading appearance of consensus“
• Falsified references: “Fortunately, in such cases, the question of AI use is irrelevant. Submitting fabricated references constitutes academic misconduct in its own right, regardless of how they were generated“
• Student confessions: Might be “confession under duress. The reasoning mistakenly assumes that because the confession follows the AI detection flag, the flag must have been correct. Correlation, however, does not imply causation. Academic integrity must rise above believing in the equivalent of horoscopes, tarot cards, or Ouija boards, simply because their predictions occasionally seem to align with real events. The confession does not retrospectively validate the AI detector’s result, just as a seemingly accurate horoscope does not prove astrology’s legitimacy” (love this quote!)
• Using generative AI, for example by generating a response by AI themselves and comparing (not proof of anything, also easy to fall into confirmation bias) or asking AI if the text was AI generated (not proof of anything, since AI cannot detect AI output).
• Past writing styles: Writing styles evolve over time, and we want students to experiment with styles to find their own voice! So this does not work either.
• Hidden adversarial prompts: You know, the white, tiny text that is supposed to catch students that copy task texts into AI and then don’t pick up on the frequent mention of “tomato” or similar: “Setting traps for students in this way relies on deception, undermines trust between students and staff, and contradicts the principles of fair assessment and academic integrity. […] The issue at the heart of this method is that it assumes dishonesty by default, treating students as inherently deceptive rather than active participants in a learning environment built on mutual respect. Universities should lead by example, upholding the same ethical standards they expect from students“

What makes things even more complicated is that text does not need to be 100% AI generated OR 100% human generated, it can exist anywhere on that spectrum, where it is generated WITH AI, and this makes AI detection tools meaningless from the start. Bassett et al. (2026) also discuss other facets of using AI detection products, like the burden of proof, the difficulty in defining what exactly is part of an assessment, security risks of uploading student texts to AI detection products, and the students’ right to silence if under investigation, which are all important considerations.

What makes this article most valuable for my work right now is however the handy list of strategies that teachers use to “validate” AI detection product’s outputs, and why they are flawed; it’s nice to have that compilation all in one place! Especially when combined with the strong conclusion: “AI detection in education is not merely flawed; it is conceptually unsound. […] Institutions must accept that AI detection is an unworkable solution to a problem that cannot be solved through surveillance and punishment. The focus must move from detection and enforcement to assessment design that recognises AI’s role in learning and the reality that unsupervised assessments cannot be secured. The continued use of AI detectors exposes students to procedural injustices and signals a fundamental misunderstanding of education’s purpose. AI detection does not safeguard academic integrity; it undermines it.”

So let’s get rid of it, and focus on human relationships!

Bassett, M. A., Bradshaw, W., Bornsztejn, H., Hogg, A., Murdoch, K., Pearce, B., & Webber, C. (2026). Heads we win, tails you lose: AI detectors in education. Journal of Higher Education Policy and Management, 1-16.

Featured image: Sjö sjön on campus today: Land-runoff (ice-runon?) nicely visible around the edges of the pond and around the island! Yes, it’s raining and melting…