A student submits an essay she wrote herself. A detector flags 90 percent of it as AI-generated. Her professor believes the tool. This isn’t a hypothetical: it’s a pattern that repeated across enough classrooms in 2023 that Vanderbilt University disabled Turnitin’s AI-detection feature entirely, doing the math on what even a small error rate meant across 75,000 annual submissions.
The uncomfortable finding underneath these stories isn’t that AI detectors are occasionally wrong. It’s that they’re wrong in a specific, predictable direction, and the people most likely to be falsely accused are often the ones least equipped to fight the accusation. The Markup’s coverage of the underlying research documents several of these cases directly.
How Do AI Detectors Actually Work?
Most AI detectors estimate the statistical predictability of text, a property called perplexity, along with how much that predictability varies sentence to sentence, called burstiness. Text that uses common word choices in a fairly uniform, low-surprise pattern scores as more likely to be machine-generated. Text with more unusual word choices and more variation in sentence complexity scores as more likely to be human.
The logic has a real basis: language models are trained to predict the statistically likely next word, so their output tends to sit closer to the “expected” pattern than natural human writing does. The problem is what else produces that same low-perplexity, low-burstiness signature – and it turns out quite a lot of ordinary human writing does.
The Research That Changed the Conversation
In 2023, a team of researchers, including several at Stanford, published a study titled “GPT detectors are biased against non-native English writers” in the journal Patterns. They ran 91 real essays written by non-native English speakers for the TOEFL exam through seven widely used AI detectors.
The results were stark. The detectors misclassified 61.3 percent of the human-written TOEFL essays as AI-generated. Nearly 98 percent of the essays were flagged by at least one detector, and almost 20 percent were unanimously flagged as AI-written by all seven tools tested. On the same detectors, essays written by native English speakers were misclassified at a rate close to zero.
Every single essay in that TOEFL sample was written by a real student, before generative AI writing tools were in common use. A survey of AI-text detection research covering this and related findings points to a plausible explanation: non-native writers tend to rely more heavily on common vocabulary and simpler, more formulaic sentence structures learned through language instruction – which is precisely the low-perplexity, low-burstiness signature that these tools are built to associate with AI output.
It’s Not Only a Non-Native-Speaker Problem
The same mechanism has been reported to affect other groups of writers whose natural style happens to overlap with what a detector considers “too predictable”:
- Writers who lean on formal, structured writing taught explicitly in school
- Neurodivergent writers, including some autistic and ADHD writers, whose prose can favor repeated phrasing and highly organized structure
- Anyone who leans heavily on grammar-checking and autocomplete tools, since heavy editing assistance can flatten a text’s natural variation
- Non-fiction and technical writers, whose vocabulary is often narrower and more repetitive by necessity
None of these groups are using AI to write. All of them can produce text that statistically resembles it, because the detector isn’t actually reading for authorship. It’s reading for a pattern, and several very human writing habits happen to produce that same pattern.
Why Detector Accuracy Claims Are Hard to Compare
Detector vendors frequently publish their own false-positive rates, and those numbers are often far lower than what independent researchers find when they test the same tools. Part of the gap comes down to what’s being tested: a detector tuned and evaluated mainly on native-speaker college essays will report a low error rate on exactly that population, while performing far worse on text it wasn’t validated against.
| Claim source | Reported false-positive rate | What was actually tested |
|---|---|---|
| Vendor-published rate (general marketing) | Often under 2–5% | Usually the vendor’s own benchmark set |
| Independent research on native-speaker essays | Roughly 5–15% | Mixed general writing samples |
| Stanford/Patterns study on non-native TOEFL essays | 61.3% | 91 real TOEFL essays, seven detectors |
The lesson isn’t that every detector is useless. It’s that a single accuracy figure quoted without saying which population it was tested on tells you very little about how the tool will behave on the specific writing sample in front of you.
What This Means If You’ve Been Flagged
If a detector has flagged something you genuinely wrote yourself, a few responses tend to be more productive than arguing with the percentage on the screen:
- Ask what specific evidence, beyond the detector score, supports the accusation. A detector output alone is a statistical guess, not proof.
- Provide your drafting history – document version history, notes, or an earlier outline – since this is concrete evidence a detector can’t fabricate or contradict.
- Point to the documented false-positive research if the accusation involves an academic or employment consequence; this is now well enough established that most institutions have policies addressing it.
- Avoid rewriting the text specifically to “beat” the detector. This can introduce the same statistical noise that has been shown to sometimes fool detectors regardless of who wrote the original, which muddies rather than resolves the actual question of authorship.
Frequently Asked Questions
Can AI detectors reliably tell if a specific document was written by a person or a model? Not with certainty. They estimate a statistical likelihood based on patterns like word predictability, and both the false-positive and false-negative rates found in independent research are high enough that no single detector score should be treated as proof.
Are AI detectors getting more accurate over time? Detection and generation are in a continuous back-and-forth: as detectors improve, newer language models produce text that’s harder to distinguish statistically from human writing, and vice versa. There’s no evidence this settles into a permanently solved state.
Why do non-native English speakers get flagged more often? Research points to a reliance on more common vocabulary and formulaic sentence structures learned in language instruction, which produces the same low-variability statistical signature that detectors associate with AI-generated text.
Does using a grammar checker or writing assistant make my writing look more like AI? It can. Heavy editing assistance tends to smooth out the natural variation in sentence structure and word choice that detectors use as a signal of human authorship, which is one reason “cyborg writing” – human drafts heavily polished by editing tools – is an area detector vendors are still investigating.
Should schools and employers stop using AI detectors entirely? That’s a genuinely contested policy question rather than a settled one. Some institutions, including Vanderbilt, have disabled detection features specifically because of the false-positive risk. Others continue using them as one input among several, generally paired with a direct conversation with the writer rather than an automatic penalty.
The Practical Takeaway
An AI detector output is a probability estimate built on statistical patterns, not a lie detector for authorship. The 2023 Stanford research remains one of the most concrete, peer-reviewed data points on how badly that estimate can misfire for specific populations of real writers, and the underlying mechanism it identified hasn’t changed. If a detector score is being used to make a consequential decision about you or someone you’re evaluating, the score alone is evidence of a statistical pattern, not evidence of what actually happened. This is also why we treat hands-on testing rather than a single automated score as central to our own review methodology, and we’ve set out our position on how we use AI internally, and where we draw the line, in our AI editorial policy. For more on why AI-generated text has the statistical signature it does in the first place, our piece on why AI models hallucinate covers the underlying prediction mechanism these detectors are trying to reverse-engineer, and our broader coverage of the field sits under artificial intelligence.