{"course":{"slug":"ai-for-students","title":"AI for Students","url":"https://addaly.com/learn/ai-for-students","level":"beginner","track":"work"},"lesson":{"slug":"the-rules-and-the-detectors","title":"The Rules, And Why Detectors Do Not Work","url":"https://addaly.com/learn/ai-for-students/the-rules-and-the-detectors","position":56,"of":73,"minutes":9,"answer":"Detectors measure how typical your writing is, not who produced it; keep a process record instead.","keyIdea":"Detectors measure how typical your writing is, not who produced it; keep a process record instead."},"body":{"format":"markdown","text":"## What your institution's policy probably says\n\nAlmost every policy lands in one of four postures:\n\n1. **Prohibited for assessed work.** Any AI use is misconduct.\n2. **Permitted with disclosure.** Use it, declare what you used it for.\n3. **Permitted at some stages only.** Commonly: fine for research, brainstorming and proofreading; not for drafting, analysis or generating content.\n4. **Left to the module leader.** Increasingly the most common, and the most confusing.\n\nIf yours is the fourth, and there is a good chance it is, then **the university's public policy page is not your answer.** The answer is in the assignment brief, and if the brief is silent, the answer is in an email you have not sent yet.\n\nSend it. \"Is it acceptable to use an AI tool to generate practice questions and to check my grammar for this essay? I want to be sure before I start.\" Ask before, not after. Keep the reply. A dated email from the person marking your work is worth more than any argument you can construct later.\n\nAlso know: policies are being rewritten constantly. A rule you learned in first year may not be the rule now. Check each term.\n\n## The detectors\n\nHere is the part nobody tells students clearly.\n\nOpenAI released an AI Text Classifier in January 2023: a detector for its own model's output. It withdrew it that July, citing low accuracy. Its own published figures: it identified about **26% of AI-written text**, and wrongly flagged about **9% of human-written text** as AI. The company that built the generator could not reliably detect the generator.\n\nA 2023 Stanford study (Liang and colleagues, published in *Patterns*) ran seven commercial detectors over TOEFL essays written by non-native English speakers, and over essays by US eighth-graders. The eighth-graders were classified correctly almost every time. **More than half of the non-native speakers' essays were flagged as AI-generated.** The reason is mechanical: the detectors keyed on low lexical variety and simple sentence construction. A careful second-language writer produces exactly that.\n\nSeveral universities responded by turning the tools off. Vanderbilt disabled Turnitin's AI detector in 2023 and said so publicly. Turnitin has claimed roughly a 1% false positive rate at document level, and even taken at face value, across twenty thousand essays in a term, one percent is two hundred students accused of something they did not do.\n\n> Figure: What the detector numbers actually are. Bars (%): AI-written text the classifier caught 26; Human-written text it wrongly flagged 9; Non-native English essays flagged as AI 61; US eighth-graders' essays flagged as AI 2. OpenAI's own published figures for its AI Text Classifier, withdrawn in July 2023 citing low accuracy: the company that built the generator could not reliably detect it. The essay figures are from Liang and colleagues in Patterns, 2023, averaged across seven commercial detectors. The mechanism explains the gap exactly. A careful writer in a second language uses common words and standard constructions, and that is what the instrument measures.\n\n## Why this is not a solvable engineering problem\n\nA detector cannot examine provenance. There is no watermark in ordinary text. What it can do is measure **statistical typicality**: how predictable each word is given the ones before it. Machine text is, on average, more predictable.\n\nSo is clear, plain, careful human writing. So is writing by someone with a smaller working vocabulary in the language of assessment. So is the prose of a student who was taught to write simply, which is what good teachers teach.\n\nEverything that makes writing easy to read makes it look machine-made. That is not a bug to be patched. It is what the measurement is.\n\n## Protect yourself before you need to\n\nBeing wrongly accused is frightening, and it is a defence built in advance or not at all.\n\n- **Write in something with revision history.** Google Docs and Word both keep timestamped version history. A document that grew over eleven sessions across two weeks looks nothing like one pasted in whole, and the history is inspectable.\n- **Keep the mess.** Your outline, your crossed-out plan, the photo of the notebook page, the draft with the abandoned second argument.\n- **Keep the sources you actually read**, with your own notes on them.\n- **Be ready to talk it through.** In a viva-style conversation, the person who wrote it can explain why they cut the third section. That is the evidence that persuades.\n\nStart the habit now. The record is worthless if you begin assembling it after the email arrives.\n\n## The thing that has to be said\n\nDetectors failing is not permission.\n\nThe argument for doing your own work is not \"you will be caught,\" because increasingly you will not be. The argument is the one from lesson two: the exam hall exists, the professional situation exists, and both are populated by people who can tell in ten minutes of conversation whether you know your subject. The detector was never the thing that was going to find out."},"check":{"question":"A detector reports your essay as \"82% AI\". You wrote every word yourself. What does that number actually mean?","options":[{"text":"It means the tool is 82% confident, so there is about a one-in-five chance you are innocent and you now have to prove it.","correct":false,"tempting":"Reads the score as a calibrated probability of guilt, which is how such numbers get used in hearings and is not what they measure."},{"text":"It means 82% of your sentences matched AI-generated text held in a database, so producing your sources will clear you.","correct":false,"tempting":"Imagines a plagiarism-style corpus match; detectors compare nothing, they score statistical predictability of your own words."},{"text":"It is a statement about how statistically typical your sentences are, not about how they were produced, and plain, clear writing scores as typical.","correct":true},{"text":"It means your writing has been unconsciously shaped by reading a lot of AI text, which is what such tools are picking up.","correct":false,"tempting":"Sounds sophisticated and even generous, but the tool has no way to distinguish stylistic influence from anything else it is measuring."}]},"prev":{"slug":"group-work-and-shared-rules","title":"Group work, and shared rules","url":"https://addaly.com/learn/ai-for-students/group-work-and-shared-rules"},"next":{"slug":"reading-your-actual-policy","title":"Reading your actual policy","url":"https://addaly.com/learn/ai-for-students/reading-your-actual-policy"},"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) 'The Rules, And Why Detectors Do Not Work', in AI for Students. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-students/the-rules-and-the-detectors (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":"The Rules, And Why Detectors Do Not Work","courseTitle":"AI for Students","module":{"index":7,"slug":"rules-honesty-record","title":"Rules, honesty and the record"},"answer":null,"keyIdea":"Detectors measure how typical your writing is, not who produced it; keep a process record instead.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 7.1","figure":{"type":"bars","unit":"%","title":"What the detector numbers actually are","series":[{"label":"AI-written text the classifier caught","value":26},{"label":"Human-written text it wrongly flagged","value":9},{"label":"Non-native English essays flagged as AI","value":61,"highlight":true},{"label":"US eighth-graders’ essays flagged as AI","value":2}],"caption":"OpenAI’s own published figures for its AI Text Classifier, withdrawn in July 2023 citing low accuracy: the company that built the generator could not reliably detect it. The essay figures are from Liang and colleagues in Patterns, 2023, averaged across seven commercial detectors. The mechanism explains the gap exactly. A careful writer in a second language uses common words and standard constructions, and that is what the instrument measures."},"alt":"Figure: What the detector numbers actually are. Bars (%): AI-written text the classifier caught 26; Human-written text it wrongly flagged 9; Non-native English essays flagged as AI 61; US eighth-graders’ essays flagged as AI 2. OpenAI’s own published figures for its AI Text Classifier, withdrawn in July 2023 citing low accuracy: the company that built the generator could not reliably detect it. The essay figures are from Liang and colleagues in Patterns, 2023, averaged across seven commercial detectors. The mechanism explains the gap exactly. A careful writer in a second language uses common words and standard constructions, and that is what the instrument measures."}},"blocks":[{"t":"h2","id":"what-your-institution-s-policy-probably-says","runs":[{"text":"What your institution’s policy probably says"}]},{"t":"p","runs":[{"text":"Almost every policy lands in one of four postures:"}]},{"t":"ol","start":1,"items":[[{"text":"Prohibited for assessed work.","b":true},{"text":" Any AI use is misconduct."}],[{"text":"Permitted with disclosure.","b":true},{"text":" Use it, declare what you used it for."}],[{"text":"Permitted at some stages only.","b":true},{"text":" Commonly: fine for research, brainstorming and proofreading; not for drafting, analysis or generating content."}],[{"text":"Left to the module leader.","b":true},{"text":" Increasingly the most common, and the most confusing."}]]},{"t":"p","runs":[{"text":"If yours is the fourth, and there is a good chance it is, then "},{"text":"the university’s public policy page is not your answer.","b":true},{"text":" The answer is in the assignment brief, and if the brief is silent, the answer is in an email you have not sent yet."}]},{"t":"p","runs":[{"text":"Send it. “Is it acceptable to use an AI tool to generate practice questions and to check my grammar for this essay? I want to be sure before I start.” Ask before, not after. Keep the reply. A dated email from the person marking your work is worth more than any argument you can construct later."}]},{"t":"p","runs":[{"text":"Also know: policies are being rewritten constantly. A rule you learned in first year may not be the rule now. Check each term."}]},{"t":"h2","id":"the-detectors","runs":[{"text":"The detectors"}]},{"t":"p","runs":[{"text":"Here is the part nobody tells students clearly."}]},{"t":"p","runs":[{"text":"OpenAI released an AI Text Classifier in January 2023: a detector for its own model’s output. It withdrew it that July, citing low accuracy. Its own published figures: it identified about "},{"text":"26% of AI-written text","b":true},{"text":", and wrongly flagged about "},{"text":"9% of human-written text","b":true},{"text":" as AI. The company that built the generator could not reliably detect the generator."}]},{"t":"p","runs":[{"text":"A 2023 Stanford study (Liang and colleagues, published in "},{"text":"Patterns","i":true},{"text":") ran seven commercial detectors over TOEFL essays written by non-native English speakers, and over essays by US eighth-graders. The eighth-graders were classified correctly almost every time. "},{"text":"More than half of the non-native speakers’ essays were flagged as AI-generated.","b":true},{"text":" The reason is mechanical: the detectors keyed on low lexical variety and simple sentence construction. A careful second-language writer produces exactly that."}]},{"t":"p","runs":[{"text":"Several universities responded by turning the tools off. Vanderbilt disabled Turnitin’s AI detector in 2023 and said so publicly. Turnitin has claimed roughly a 1% false positive rate at document level, and even taken at face value, across twenty thousand essays in a term, one percent is two hundred students accused of something they did not do."}]},{"t":"h2","id":"why-this-is-not-a-solvable-engineering-problem","runs":[{"text":"Why this is not a solvable engineering problem"}]},{"t":"p","runs":[{"text":"A detector cannot examine provenance. There is no watermark in ordinary text. What it can do is measure "},{"text":"statistical typicality","b":true},{"text":": how predictable each word is given the ones before it. Machine text is, on average, more predictable."}]},{"t":"p","runs":[{"text":"So is clear, plain, careful human writing. So is writing by someone with a smaller working vocabulary in the language of assessment. So is the prose of a student who was taught to write simply, which is what good teachers teach."}]},{"t":"p","runs":[{"text":"Everything that makes writing easy to read makes it look machine-made. That is not a bug to be patched. It is what the measurement is."}]},{"t":"h2","id":"protect-yourself-before-you-need-to","runs":[{"text":"Protect yourself before you need to"}]},{"t":"p","runs":[{"text":"Being wrongly accused is frightening, and it is a defence built in advance or not at all."}]},{"t":"ul","items":[[{"text":"Write in something with revision history.","b":true},{"text":" Google Docs and Word both keep timestamped version history. A document that grew over eleven sessions across two weeks looks nothing like one pasted in whole, and the history is inspectable."}],[{"text":"Keep the mess.","b":true},{"text":" Your outline, your crossed-out plan, the photo of the notebook page, the draft with the abandoned second argument."}],[{"text":"Keep the sources you actually read","b":true},{"text":", with your own notes on them."}],[{"text":"Be ready to talk it through.","b":true},{"text":" In a viva-style conversation, the person who wrote it can explain why they cut the third section. That is the evidence that persuades."}]]},{"t":"p","runs":[{"text":"Start the habit now. The record is worthless if you begin assembling it after the email arrives."}]},{"t":"h2","id":"the-thing-that-has-to-be-said","runs":[{"text":"The thing that has to be said"}]},{"t":"p","runs":[{"text":"Detectors failing is not permission."}]},{"t":"p","runs":[{"text":"The argument for doing your own work is not “you will be caught,” because increasingly you will not be. The argument is the one from lesson two: the exam hall exists, the professional situation exists, and both are populated by people who can tell in ten minutes of conversation whether you know your subject. The detector was never the thing that was going to find out."}]}],"check":{"question":[{"text":"A detector reports your essay as “82% AI”. You wrote every word yourself. What does that number actually mean?"}],"options":[{"text":[{"text":"It means the tool is 82% confident, so there is about a one-in-five chance you are innocent and you now have to prove it."}],"correct":false,"tempting":[{"text":"Reads the score as a calibrated probability of guilt, which is how such numbers get used in hearings and is not what they measure."}]},{"text":[{"text":"It means 82% of your sentences matched AI-generated text held in a database, so producing your sources will clear you."}],"correct":false,"tempting":[{"text":"Imagines a plagiarism-style corpus match; detectors compare nothing, they score statistical predictability of your own words."}]},{"text":[{"text":"It is a statement about how statistically typical your sentences are, not about how they were produced, and plain, clear writing scores as typical."}],"correct":true,"tempting":null},{"text":[{"text":"It means your writing has been unconsciously shaped by reading a lot of AI text, which is what such tools are picking up."}],"correct":false,"tempting":[{"text":"Sounds sophisticated and even generous, but the tool has no way to distinguish stylistic influence from anything else it is measuring."}]}],"order":[0,2,1,3],"why":null},"prev":{"slug":"group-work-and-shared-rules","title":"Group work, and shared rules","minutes":7},"next":{"slug":"reading-your-actual-policy","title":"Reading your actual policy","minutes":8}}}