{"course":{"slug":"ai-for-teachers","title":"AI for Teachers","url":"https://addaly.com/learn/ai-for-teachers","level":"beginner","track":"work"},"lesson":{"slug":"detecting-ai-writing","title":"Why detectors fail, and what to do instead","url":"https://addaly.com/learn/ai-for-teachers/detecting-ai-writing","position":34,"of":57,"minutes":9,"answer":"Detectors flag careful second-language writing as AI; use process evidence and conversation instead.","keyIdea":"Detectors flag careful second-language writing as AI; use process evidence and conversation instead."},"body":{"format":"markdown","text":"## How a detector claims to work\n\nIt looks at two statistical properties. How predictable each word is given the words before it, and how much sentence structure varies across the piece. AI text tends to be more predictable and more even.\n\nSo does the writing of a careful student working in their second or third language, who chooses safe vocabulary and keeps sentences to a familiar shape.\n\nThat is the entire problem, in one sentence. The signal the detector reads is not \"machine\", it is \"unsurprising\", and plenty of human writing is unsurprising for reasons that have nothing to do with cheating.\n\n## What the evidence actually shows\n\nA 2023 study published in *Patterns* tested seven detectors on TOEFL essays written by non-native English speakers. More than half were flagged as AI-generated. Essays by US school students were almost all correctly cleared. When the same non-native essays were rewritten with richer vocabulary, the flags largely disappeared, which tells you exactly what was being measured.\n\nOpenAI withdrew its own detection tool in July 2023, citing a low rate of accuracy. The company with the most access to how these models write could not build a reliable detector for their output.\n\nVendors publish low false-positive rates from their own testing. Independent evaluations have generally found higher rates, and higher still for second-language writers. Treat vendor figures as a claim, not a finding.\n\n## The arithmetic nobody does\n\nSuppose a detector does have a 1% false positive rate, better than most claims. You have 30 students. Run every piece of homework through it and you can expect a wrong flag against an innocent student roughly one class in three, week after week.\n\nMeanwhile, running AI text through a free paraphrasing tool defeats most detectors in about eight seconds.\n\nSo the tool punishes the honest and misses the determined. That is the worst possible error profile, and no improvement in accuracy fixes its direction.\n\n## What to do instead\n\n**Collect process, not just product.** Ask for the draft, the notes, the plan. Word and Google Docs both keep version history for free, and it is already on. A document that appeared in one paste at 11pm looks different from one that grew over four days.\n\n**Have the two-minute conversation, with everyone.** \"Talk me through why you chose this example.\" Do it with every student for every major piece and it is a normal part of the course, not an accusation. Do it with one student and it is an interrogation. The difference is entirely in whether it is routine.\n\n**Get a baseline.** One piece of in-class writing early in the term, on paper, tells you what each student sounds like. Not as evidence to convict: as the thing that makes you notice, honestly, when something reads oddly.\n\n> Figure: One hundred pieces through a detector claiming 1% false positives. A two by two grid, rows Detector flags it and Detector clears it, columns Written with AI and Written by the student: Detector flags it and Written with AI: 3; Detector flags it and Written by the student: 4; Detector clears it and Written with AI: 5; Detector clears it and Written by the student: 88. Eight of these hundred pieces were written with AI. Three were caught, five were missed, and four honest students were flagged. Fewer than half the flags are right, and the wrong ones land on the students whose plain, careful prose reads as unsurprising. Running the missed five through a free paraphraser takes eight seconds.\n\n**Write the policy before you need it.** Students should know what is permitted before they sit down, not after you are unhappy.\n\n## If you are worried about a specific piece\n\nDo not open with the accusation, and never with a score. A detector percentage is not evidence and cannot be shown to a parent.\n\nOpen with curiosity: \"This is interesting. Tell me how you got to this argument.\" A student who wrote it will talk happily for two minutes. A student who did not will usually tell you, or make it obvious without you having to say the word.\n\nAnd hold the line internally. If a colleague or a head wants to act on a detector score alone, the answer is that the tool has a known bias against second-language writers, its own makers withdrew theirs for inaccuracy, and a wrong accusation costs a child far more than a missed one costs the school."},"check":{"question":"A school buys a detector advertised at 98% accuracy. The head says that is good enough to act on. Why is a teacher right to push back?","options":[{"text":"The 2% does not fall evenly (the students wrongly flagged are disproportionately second-language and plainer writers), and the cost of a wrong accusation to a child far exceeds the cost of missing one cheat.","correct":true},{"text":"98% is simply not high enough to act on; a detector at 99.9% would be usable.","correct":false,"tempting":"Frames it as a threshold that better technology will cross, when the failures are concentrated on particular students rather than scattered."},{"text":"The 98% comes from the vendor's own testing, and independent evaluation would find it lower.","correct":false,"tempting":"True as far as it goes, and it lets you keep believing that an honestly measured 98% would be fine to act on."},{"text":"Detectors only recognise the models they were trained against, so they will stop working as new AI arrives.","correct":false,"tempting":"A real limitation, but it argues the tool will get worse rather than that acting on it was never fair."}]},"prev":{"slug":"peer-assessment-with-a-third-marker","title":"Peer assessment with a third marker","url":"https://addaly.com/learn/ai-for-teachers/peer-assessment-with-a-third-marker"},"next":{"slug":"assignment-redesign","title":"Assignments AI cannot do for you","url":"https://addaly.com/learn/ai-for-teachers/assignment-redesign"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Why detectors fail, and what to do instead', in AI for Teachers. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-teachers/detecting-ai-writing (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Why detectors fail, and what to do instead","courseTitle":"AI for Teachers","module":{"index":5,"slug":"integrity-and-ai-written-work","title":"Integrity: AI-written work and what to do about it"},"answer":null,"keyIdea":"Detectors flag careful second-language writing as AI; use process evidence and conversation instead.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 5.1","figure":{"cols":["Written with AI","Written by the student"],"rows":["Detector flags it","Detector clears it"],"type":"matrix","cells":[["3","4"],["5","88"]],"title":"One hundred pieces through a detector claiming 1% false positives","caption":"Eight of these hundred pieces were written with AI. Three were caught, five were missed, and four honest students were flagged. Fewer than half the flags are right, and the wrong ones land on the students whose plain, careful prose reads as unsurprising. Running the missed five through a free paraphraser takes eight seconds.","cellLabels":[["caught","an accusation against an honest student"],["missed","correctly cleared"]]},"alt":"Figure: One hundred pieces through a detector claiming 1% false positives. A two by two grid, rows Detector flags it and Detector clears it, columns Written with AI and Written by the student: Detector flags it and Written with AI: 3; Detector flags it and Written by the student: 4; Detector clears it and Written with AI: 5; Detector clears it and Written by the student: 88. Eight of these hundred pieces were written with AI. Three were caught, five were missed, and four honest students were flagged. Fewer than half the flags are right, and the wrong ones land on the students whose plain, careful prose reads as unsurprising. Running the missed five through a free paraphraser takes eight seconds."}},"blocks":[{"t":"h2","id":"how-a-detector-claims-to-work","runs":[{"text":"How a detector claims to work"}]},{"t":"p","runs":[{"text":"It looks at two statistical properties. How predictable each word is given the words before it, and how much sentence structure varies across the piece. AI text tends to be more predictable and more even."}]},{"t":"p","runs":[{"text":"So does the writing of a careful student working in their second or third language, who chooses safe vocabulary and keeps sentences to a familiar shape."}]},{"t":"p","runs":[{"text":"That is the entire problem, in one sentence. The signal the detector reads is not “machine”, it is “unsurprising”, and plenty of human writing is unsurprising for reasons that have nothing to do with cheating."}]},{"t":"h2","id":"what-the-evidence-actually-shows","runs":[{"text":"What the evidence actually shows"}]},{"t":"p","runs":[{"text":"A 2023 study published in "},{"text":"Patterns","i":true},{"text":" tested seven detectors on TOEFL essays written by non-native English speakers. More than half were flagged as AI-generated. Essays by US school students were almost all correctly cleared. When the same non-native essays were rewritten with richer vocabulary, the flags largely disappeared, which tells you exactly what was being measured."}]},{"t":"p","runs":[{"text":"OpenAI withdrew its own detection tool in July 2023, citing a low rate of accuracy. The company with the most access to how these models write could not build a reliable detector for their output."}]},{"t":"p","runs":[{"text":"Vendors publish low false-positive rates from their own testing. Independent evaluations have generally found higher rates, and higher still for second-language writers. Treat vendor figures as a claim, not a finding."}]},{"t":"h2","id":"the-arithmetic-nobody-does","runs":[{"text":"The arithmetic nobody does"}]},{"t":"p","runs":[{"text":"Suppose a detector does have a 1% false positive rate, better than most claims. You have 30 students. Run every piece of homework through it and you can expect a wrong flag against an innocent student roughly one class in three, week after week."}]},{"t":"p","runs":[{"text":"Meanwhile, running AI text through a free paraphrasing tool defeats most detectors in about eight seconds."}]},{"t":"p","runs":[{"text":"So the tool punishes the honest and misses the determined. That is the worst possible error profile, and no improvement in accuracy fixes its direction."}]},{"t":"h2","id":"what-to-do-instead","runs":[{"text":"What to do instead"}]},{"t":"p","runs":[{"text":"Collect process, not just product.","b":true},{"text":" Ask for the draft, the notes, the plan. Word and Google Docs both keep version history for free, and it is already on. A document that appeared in one paste at 11pm looks different from one that grew over four days."}]},{"t":"p","runs":[{"text":"Have the two-minute conversation, with everyone.","b":true},{"text":" “Talk me through why you chose this example.” Do it with every student for every major piece and it is a normal part of the course, not an accusation. Do it with one student and it is an interrogation. The difference is entirely in whether it is routine."}]},{"t":"p","runs":[{"text":"Get a baseline.","b":true},{"text":" One piece of in-class writing early in the term, on paper, tells you what each student sounds like. Not as evidence to convict: as the thing that makes you notice, honestly, when something reads oddly."}]},{"t":"p","runs":[{"text":"Write the policy before you need it.","b":true},{"text":" Students should know what is permitted before they sit down, not after you are unhappy."}]},{"t":"h2","id":"if-you-are-worried-about-a-specific-piece","runs":[{"text":"If you are worried about a specific piece"}]},{"t":"p","runs":[{"text":"Do not open with the accusation, and never with a score. A detector percentage is not evidence and cannot be shown to a parent."}]},{"t":"p","runs":[{"text":"Open with curiosity: “This is interesting. Tell me how you got to this argument.” A student who wrote it will talk happily for two minutes. A student who did not will usually tell you, or make it obvious without you having to say the word."}]},{"t":"p","runs":[{"text":"And hold the line internally. If a colleague or a head wants to act on a detector score alone, the answer is that the tool has a known bias against second-language writers, its own makers withdrew theirs for inaccuracy, and a wrong accusation costs a child far more than a missed one costs the school."}]}],"check":{"question":[{"text":"A school buys a detector advertised at 98% accuracy. The head says that is good enough to act on. Why is a teacher right to push back?"}],"options":[{"text":[{"text":"The 2% does not fall evenly (the students wrongly flagged are disproportionately second-language and plainer writers), and the cost of a wrong accusation to a child far exceeds the cost of missing one cheat."}],"correct":true,"tempting":null},{"text":[{"text":"98% is simply not high enough to act on; a detector at 99.9% would be usable."}],"correct":false,"tempting":[{"text":"Frames it as a threshold that better technology will cross, when the failures are concentrated on particular students rather than scattered."}]},{"text":[{"text":"The 98% comes from the vendor’s own testing, and independent evaluation would find it lower."}],"correct":false,"tempting":[{"text":"True as far as it goes, and it lets you keep believing that an honestly measured 98% would be fine to act on."}]},{"text":[{"text":"Detectors only recognise the models they were trained against, so they will stop working as new AI arrives."}],"correct":false,"tempting":[{"text":"A real limitation, but it argues the tool will get worse rather than that acting on it was never fair."}]}],"order":[3,1,0,2],"why":null},"prev":{"slug":"peer-assessment-with-a-third-marker","title":"Peer assessment with a third marker","minutes":8},"next":{"slug":"assignment-redesign","title":"Assignments AI cannot do for you","minutes":8}}}