Software that claims to spot AI writing is now sold to schools and universities everywhere. This walks through what those tools actually do, then shows the number trap that makes their scores so easy to misread. A live calculator lets you watch honest students pile up in the flagged column. The point is not that detectors are useless; it is that a score is a guess about text, never a measurement of who wrote it.
Module 01
What an AI detector actually is
In this module: the four techniques, and the one that education tools nearly all use
Before any number means anything, it helps to know what the tool did to produce it. Detection is not one method. It is a label over several, and the one you meet at work has a built-in limit.
One label, four techniques
A 2025 survey in Computational Linguistics sorts the field into four families.
WatermarkingThe provider marks the text as it is generated.A statistical signal woven in at the source. It only works if the AI company chose to add it, and it can wash out when the text is edited.
StatisticalMeasure the maths of the text, with no training.Look at how predictable the words are. This is the perplexity and burstiness idea on the next slide.
Supervised classifierA model trained to tell two piles of text apart.Trained on labelled human and AI writing. Turnitin, GPTZero, Pangram and their rivals are nearly all this kind.
Human-assistedThe software supports a person's judgement.The tool flags things for a human to weigh, rather than returning a verdict on its own.
The detectors staff actually meet in education are almost all supervised classifiers. That choice sets the limit worth sitting with.
Perplexity and burstiness
A first-generation detector runs on two measures. Perplexity asks how surprising each word is to a language model; AI text tends to pick likely words, so it reads as smooth and predictable. Burstiness asks how much that surprise varies; human writing lurches between plain and odd sentences, while machine text stays even. Low perplexity plus low burstiness reads as “AI”.
Put plainly, the detector runs a next-word prediction machine in reverse over an essay, and asks how often the writer picked the boring, expected word. That is a real signal. It is also one that plenty of honest writing sets off.
What the statistics get wrong
Guess first. Which famous document do AI detectors keep flagging as machine-written?
The United States Declaration of Independence. The model has effectively memorised it, so it scores as highly predictable, which the tool reads as AI. Detectors have flagged the US Constitution and passages of the Bible the same way. The deeper problem is who else writes in a predictable register. That includes formulaic and templated work, much competent student writing, most VET assessment answers, and people writing in English as an additional language.In a 2023 Patterns study, seven detectors wrongly flagged an average of 61 per cent of essays by non-native English writers as AI, while being near-perfect on native writers. Asking the AI to “elevate the language” flipped the result, which shows the tool is reading vocabulary, not authorship.
Inside a modern detector
A supervised classifier has the same shape whatever brand is on the report.
That last line is the whole module. The output is a model’s guess from patterns it was trained on, not a reading of who held the pen.
Module 02
The base-rate trap
In this module: why a 92 per cent score is not a 92 per cent chance of cheating
This is the part that does the damage in real integrity meetings. The score on the report and the chance a flagged student cheated are two different numbers, and the gap between them is wide.
The misreading
A report says 92% AI. A tired marker reads that as a 92 per cent chance this student used AI. That is the wrong reading. A detector’s error rates are properties of the test. The chance a flagged student actually cheated is a different quantity, and it depends on something the score cannot contain: how common real misuse is in this cohort.
Australian research (Bassett and colleagues, 2026) found this misreading is common among the staff who make integrity decisions. The fix is not a better detector; it is understanding base rates.
Same tool, two very different answers
A detector catches 90 per cent of AI work and wrongly flags just 1 per cent of honest work. In a class where only 1 student in 100 truly misuses AI, what share of all the flags are false accusations?
About half. The small 1 per cent error is applied to the large honest majority, and those false flags rival the real ones. If instead 5 in 100 genuinely misuse AI, about 1 flag in 6 is false, from the very same tool with the very same advertised accuracy. The number that changed was not the detector; it was the base rate.Bassett and colleagues (2026) run this arithmetic. They find a flagged paper’s true chance of being AI can range from about 97 per cent down to under 48 per cent. It depends entirely on the cohort’s real misuse rate, which nobody can measure.
Run the numbers yourself
Set the cohort, how common misuse really is, and the detector’s two error rates. Watch the flagged column. The misuse slider is the honest unknown; no one knows its true value for a real class.
Base-rate calculator: who ends up flagged?
Flags on genuine misuseFlags on honest work
55
Total flags raised
10
Honest students flagged
83%
Chance a given flag is genuine
At these settings, about 1 flag in every 6 lands on a student who did nothing wrong.
What that looks like at scale
When Vanderbilt University turned off Turnitin’s AI detector in August 2023, the reason was arithmetic. A claimed 1 per cent false-positive rate, across roughly 75,000 submissions a year, meant about 750 honest papers flagged annually. That is 750 integrity conversations that should never have started.
The calculator is not a party trick. It is the same sum a large university did before pulling the tool.
Module 03
What a score can and cannot establish
In this module: why headline accuracy fails on real work, and why an all-clear proves nothing
Vendors quote high accuracy, and it is usually true of something: a chosen benchmark, at a moment, on text that is wholly human or wholly AI. Real coursework is none of those things.
Where the headline numbers fall apart
Fair settingCatch rate collapses when false flags are capped.Held to a limit of 1 honest document flagged in 100, seven detectors caught between 58 and 3 per cent of AI text, and zero on some tasks. NAACL, 2025.
Light evasionSmall changes cut accuracy hard.The RAID benchmark found simple settings changes cut accuracy by up to 38 points; synonym and character tricks cut 36 to 41. ACL, 2024.
Low skillBasic tricks are enough.Paraphrasing and added spelling errors dropped accuracy by an average of 17 per cent. Perkins and colleagues, 2024.
Real textWeakest exactly where student work lives.On edited, mixed, or short text, reliability falls away. One 2026 test put Turnitin at 0.61 accuracy, with almost no detection of hybrid human-and-AI writing.
These are peer-reviewed and independent findings, not a rival vendor’s marketing. Most contested submissions are exactly the mixed, edited, short kind.
The traffic runs both ways
A flag is not proof of AI use. An all-clear is not proof of none. Independent testing finds detectors lean toward calling text human. Tools are sold specifically to launder AI writing past them. And a 2026 study pushed a leading detector to rate machine output “100 per cent human” through repeated paraphrasing. Neither direction of the result carries the weight people put on it.
Even the strongest independent result comes fenced. A 2025 University of Chicago paper found the specialist detector Pangram near error-free. That held only on longer passages that were wholly human or wholly AI; short-passage false flags still climbed toward 1 per cent. That is not the shape of most real submissions.
Why this matters
None of this says detection is worthless. It says a detector score is a lead, not a verdict. The score cannot be turned into a probability of cheating without a number nobody has: the cohort’s real rate of misuse. Treat a flag as a reason to look closer, through a fair process and other evidence, never as the finding itself. And remember who the statistics fall hardest on: the templated writer, the second-language writer, the student whose honest work simply reads as predictable.
Sources used. Detector taxonomy and out-of-distribution failure: “A Survey on LLM-Generated Text Detection”, Computational Linguistics 51:1, 2025. Non-native-writer bias: Liang et al., “GPT detectors are biased against non-native English writers”, Patterns, 10 July 2023 (United States data). Catch rate at a fixed 1 per cent false-positive rate: Tufts et al., NAACL 2025. Adversarial fragility: Dugan et al., “RAID”, ACL 2024. Low-skill evasion: Perkins, Roe, Vu et al., International Journal of Educational Technology in Higher Education, 9 September 2024. Accuracy on hybrid and edited text: Hadra, Cambridge and Mesbah, International Journal for Educational Integrity, February 2026; Weber-Wulff et al., same journal, December 2023. The base-rate worked example and staff misreadings: Bassett et al., “Heads we win, tails you lose”, Journal of Higher Education Policy and Management, online 29 January 2026 (Australian authors). Strongest independent detector result, and its limits: Jabarian and Imas, University of Chicago BFI working paper, 2025 (not yet peer reviewed). Reliability limits on edited, mixed and short text: Guy Curtis, TEQSA Academic Integrity Toolkit. The 750-papers calculation: Vanderbilt University, 16 August 2023. Turnitin’s own error-rate statements are vendor claims, not independent validation. Checked August 2026.