A widely repeated story claims that a Danish, an Australian and a Chilean high school jointly abandoned written homework after AI text detectors failed. No source supports it. What is documented is different: detectors are wrong often enough to collapse under a disciplinary challenge, Denmark bans AI in exams rather than abolishing them, and it is individual teachers — not school systems — moving writing back into the classroom.
The story circulates in this shape: three high schools — one in Denmark, one in Australia, one in Chile — stopped assigning written homework and rebuilt their assessment from scratch after concluding that AI text detectors don’t work. It is a well-built story, with just the right amount of geographic exoticism, and it has one flaw: it cannot be found. No schools, no board decisions, no coordination between three education systems that have nothing in common. If you are looking for that template to copy, you are looking for something nobody published.
What does exist, and is verifiable, is the first half of it: the software doesn’t work well enough. We are talking about a specific class of system — machine learning classifiers trained to decide, from a text alone, whether a person or a language model wrote it. They don’t read for meaning and know nothing about the student. They measure statistical properties of the text’s surface: how predictable the next word is, how much sentence length varies, how closely the prose resembles what a large model produces on average. Out of those measures comes a number, and in the hands of a disciplinary panel that number becomes an accusation. That is where the chain breaks.
What the tests on detectors actually show
The strongest reference is the 2023 study coordinated by Debora Weber-Wulff: fourteen detection tools put through their paces, including Turnitin and GPTZero. None exceeded 80% accuracy; only five cleared 70%. Two details matter more than the headline figure. First, these tools tend to classify texts as human-written, which means most of their error doesn’t punish cheaters — it makes the tool useless as a filter. Second, accuracy degrades when text is paraphrased, which is precisely the scenario involving a student motivated not to get caught.
On vendor-reported numbers the picture is openly contradictory, and it is worth saying so rather than picking the convenient figure. Turnitin states a false positive rate below 1% and a false negative rate of 15%. Independent testing by the Washington Post produced far higher false positive rates, up to 50%, but on a small sample. Both figures have a problem: one comes from the company selling the product, the other from a sample too small to generalise. Anyone telling you detectors are 98% accurate, and anyone telling you they never work, is going beyond what the sources support.
Then there is the point that should be decisive for a school: the error is not randomly distributed. Detectors disproportionately penalise people writing in a language that is not their first, and neurodivergent writers. Plain prose with regular constructions and a limited vocabulary statistically resembles generated prose. Not because it is, but because the classifier measures predictability — and predictability is also what someone still learning the language produces.
Even the model builders backed off. On 1 February 2023 OpenAI released a classifier for AI-written text, describing it in the same announcement as “not fully reliable”, and withdrew it that July for poor accuracy. When the maker of the model cannot reliably recognise its own model’s output, the claim that a third-party vendor can do so with forensic precision becomes hard to defend.
Denmark is doing the opposite of what it’s credited with
In the romanticised version, Denmark is the country that gave up on written exams. In the reality documented by the Danish education ministry, it made the mirror-image choice: generative AI tools are banned in all on-site written exams in the folkeskole, the compulsory school covering roughly ages 6 to 16. The ban covers the written Danish and spelling papers of the ninth-year exam, mathematics, and the Danish, German, French and English papers of the FP10 track.
The interesting detail lies elsewhere, and for anyone drafting a policy it is worth more than the ban itself. In the ninth-year written Danish exam and in two parts of FP10, the internet remains allowed: students may look things up online. What is prohibited is accessing generative AI tools and quoting their output. That distinction separates consultation from production, and it can be written on a sheet of exam rules without buying any software. For oral exams based on materials produced during lessons, the decision rests with the individual school board, with the STUK agency recommending against allowing generative AI in exam products. In adult general education, basic adult education and vocational training, using ChatGPT in exams is classified as exam fraud and results in exclusion from the test.
As for Australia, the state of New South Wales banned ChatGPT in public schools in early 2023 and later built its own assistant in-house, NSWEduChat, piloted from February 2024 in a little over a dozen schools. That is a decision about which tool to supply, not about assessment. In Chile, the only documented episode is a rubber stamp used by one teacher — “suspected AI use, must explain orally” — which went viral in July 2025: a personal initiative, not school policy.
The real change is classroom-sized
The trend the invented story was reaching for does exist, but it is made of individual teachers, nearly all in the United States, not national systems. Tristen Chang, who teaches humanities to a sixth-grade class in San Francisco, scrapped take-home essays and replaced them with oral exams built around fifteen questions, some lasting five minutes. Bobson Wong, a mathematics teacher in New York, cut back on homework problems, moved practice into class time, and added questions about the method a student chose rather than the answer they reached.
These are tiny interventions, which is exactly why they replicate. They share one property: they don’t try to determine whether a text was generated. They put the student in a situation where the question doesn’t arise. A five-minute oral is not a better detector — it is a different task, one in which a student who outsourced the writing has nothing to say about the method.
What can be done tomorrow
- Separate consultation from production in the rules, as the Danish papers do: online search can stay legitimate where generative AI is not.
- Never use a classifier’s output as evidence in a disciplinary proceeding. With accuracy below 80% and error skewed against second-language and neurodivergent students, it will not hold.
- Add a short oral component to written work — even a few minutes — instead of abolishing the written task.
- Move the writing that is meant to be assessed into the classroom, and leave at home the writing meant for practice.
One last thing that barely enters the debate. A 2025 RAND survey indicates that more than half of US students, and of teachers of English, mathematics and science, use artificial intelligence for schoolwork. If both sides of the desk are using it, the detector isn’t policing a boundary: it is drawing one where it no longer exists, and drawing it badly. In 2025 a Harvard randomised controlled trial found learning gains with an AI tutor, while a small MIT Media Lab study measured lower cognitive effort in people writing essays with ChatGPT. The two findings don’t contradict each other: they say what matters is what you ask the tool to do. Which is, precisely, a teaching decision — not a procurement one.