Assessment and Academic Integrity in the Age of AI

AI detection tools output probabilities, not proof. They estimate how closely a text resembles typical machine-generated writing, and they are wrong often enough to matter. False positives are not an edge case: careful, formulaic, or heavily revised human writing can score as AI-generated, and research from Stanford in 2023 found that detectors disproportionately flagged the work of non-native English speakers, whose prose tends toward the safer, more regular constructions the tools associate with machines. OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy. When the vendor of the most widely used AI models declines to stand behind detection, institutions should take note. The harm is asymmetric. A missed case of AI misuse is a problem; a false accusation is a catastrophe for the student concerned, and the students most likely to be wrongly flagged are often those least equipped to defend themselves. In the UK, the Joint Council for Qualifications treats AI misuse in assessments as malpractice, but its guidance is clear that a detector score alone is not a sufficient basis for action. An allegation needs evidence about the work and the student, not a percentage from a tool whose workings neither the teacher nor the student can inspect. Detection is also a losing arms race. Paraphrasing tools, so-called humanisers, and each new generation of models erode whatever signal detectors briefly held. A strategy built on catching AI text is a strategy that degrades every year. None of this means integrity does not matter; it means the weight has to be carried somewhere else.