Institutions responding to generated coursework bought detection software, and most have since restricted or abandoned it. The reason is a mismatch between what the tools output and what a disciplinary process requires.
The tools measure predictability, not authorship
Detectors score how statistically typical a text is, on the basis that generated writing sits closer to the average than human writing.
That is a real signal and it is a property of the prose, not a record of how it was produced.
Writing that is clear, formulaic or heavily edited scores as machine-like, and idiosyncratic generated text scores as human.
False accusations concentrate on particular students
Text by writers working in a second language is flagged disproportionately, because a more restricted vocabulary and simpler constructions produce exactly the predictability the detector measures.
Students following a taught structure closely, and students using accessibility tools that regularise their prose, are affected in the same way.
The errors are therefore not randomly distributed, which turns a technical limitation into a fairness problem.
A probability cannot support a finding
Academic misconduct processes require evidence a student can answer.
A score stating that a passage is probably machine-generated offers nothing to respond to, since the student cannot demonstrate the negative and the panel cannot examine the reasoning.
Institutions that proceeded on scores alone faced appeals they could not defend, which is what drove the retreat more than any technical critique.
Evasion is trivial
Paraphrasing generated text, or asking a model to write less predictably, reduces detector scores substantially.
The detector therefore penalises the careless and misses the deliberate, which inverts the intended effect of any deterrent.
Detection performance also degrades with each model generation, since newer systems produce less uniform prose than the ones the detectors were calibrated against.
Assessment design is the alternative that works
Institutions that stopped trying to detect have moved to assessment that generated text cannot satisfy: work tied to a specific seminar discussion, drafts submitted in stages, oral defence of a submitted piece.
These raise marking workload, which is the real cost and the reason detection was attractive in the first place.
Version history from a document editor has proven more useful than any detector, because it shows the process rather than judging the product, and a student can point to it rather than having to disprove a score.