Seven major AI detectors were tested against 91 TOEFL essays. Real essays, written by real students, under supervised conditions. Over 60% were flagged as AI-generated. These students were being accused of cheating by the very tools their schools trusted.
AI detectors were supposed to protect academic integrity. Instead, they're disproportionately punishing non-native English speakers, neurodivergent students, and anyone whose prose happens to be simple and consistent. The collateral damage is real and growing.
The Numbers
A Stanford-affiliated study by Liang et al. (2023) using 91 TOEFL essays found that seven major AI detectors flagged over 60% of human-written ESL essays as AI-generated. These were real essays, written by real students, under supervised conditions, yet the algorithms couldn't tell.
A larger 2025 study analyzing over 10,000 text samples found that false positive rates exceeded 20% for non-native English speakers and creative writing (MDPI Information, 2025). That means in a class of 30 multilingual students, six could be wrongly accused of cheating on any given assignment.
Meanwhile, a comparative study submitted 1,212 text samples (including ESL graduate student texts written before ChatGPT existed) to popular detection tools like Copyleaks, ZeroGPT, and Scribbr. ESL writing was disproportionately flagged as AI-generated, even when the tool had no reason to suspect AI involvement.
Why This Happens: A Built-In Flaw
The bias isn't a bug that can be patched. It's structural.
Most AI detectors rely on two statistical measures: perplexity (how predictable the word choices are) and burstiness (how much sentence length and complexity vary). AI-generated text tends to have low perplexity (predictable) and low burstiness (uniform). The problem is that non-native English speakers also tend to write with limited vocabulary, simpler sentence structures, and more uniform patterns. Not because they're using AI, but because they're writing in their second, third, or fourth language.
As UCLA's HumTech research explains, detectors confuse non-native writing with AI-generated text because essays with simpler vocabulary and repetitive structures score as "low perplexity," the same statistical signature that detectors associate with AI output. The tool cannot distinguish between a student who lacks fluency and a language model that generates fluent but generic text.
No amount of fine-tuning fixes this. The detection method itself treats "simple, consistent writing" as suspicious, and that captures a huge number of honest students.
The Human Cost
False accusations of AI cheating carry severe consequences:
- Academic penalties including failing grades, course removal, or expulsion
- Psychological harm: students report anxiety, shame, and loss of motivation after being wrongly accused
- Erosion of trust between students and teachers, undermining the educational relationship
- Disproportionate impact on already-marginalized student populations
The University of San Diego Legal Research Center highlights the unreliability of AI detection tools and the risks institutions face when relying on them for misconduct determinations. When accusations are later found to be unfounded, the damage to the student-teacher relationship is already done.
Tweaking the Algorithms Doesn't Help
Some detection tool companies have responded by refining their algorithms and adding sensitivity adjustments. While this has reduced the worst false positive rates, the underlying methodology remains probabilistic. Brandeis University's AI Teaching & Learning resources caution that detection tools remain unreliable and should not be used as the sole basis for academic integrity decisions.
Content-based detection makes statistical guesses about whether text was human- or machine-generated. No matter how sophisticated the algorithm, it cannot definitively prove authorship from text analysis alone.
Proving Who Wrote It, Not Guessing What Did
There's a different way to approach this. Instead of analyzing finished text for statistical patterns, behavioral verification tracks the process of writing itself:
- Analyzes finished text
- Statistical guessing
- 20%+ false positives for ESL
- Easily evaded by paraphrasing
- Bias against simpler writing
- Tracks the writing process
- Observable behavioral evidence
- Near-zero false positives
- Cannot fake typing patterns
- Language-agnostic
A 2024 study published in the Journal of Educational Measurement found that keystroke behavior patterns were effective at detecting nonauthentic texts in writing assessments. Another 2024 paper demonstrated that keystroke dynamics could detect AI-assisted writing with up to 85.72% accuracy, and crucially, without the bias against ESL writers that plagues content-based detectors.
The key advantage: behavioral verification doesn't care about vocabulary complexity or sentence structure. It measures how you type, not what you type. A non-native English speaker with a limited vocabulary has just as unique a typing pattern as a native speaker with a PhD in literature.
Our Recommendations
- Never rely on AI detection scores alone. Detection tools should be one input among many, not a verdict.
- Use process-based evidence. Track how students write, not just what they produce. Writing process data (drafts, revision history, keystroke logs) provides far stronger evidence than statistical text analysis.
- Have conversations first. If you suspect AI use, talk to the student about their writing process before making accusations. Ask them to explain their reasoning, walk through their sources, or continue writing on the topic.
- Be aware of bias. ESL students, neurodivergent students, and students with simpler writing styles are at higher risk of false positives. Factor this into your interpretation of detection results.
- Focus on learning. The goal isn't to catch cheaters. It's to ensure students are actually learning. Tools that support the writing process serve both purposes simultaneously.
You Can't Win an Arms Race Against Language Models
Detectors try to catch AI text. AI tools evolve to evade detection. Each new model generates more human-like text, making content-based detection progressively harder. This cycle has no endpoint, and students are the ones paying for it.
Behavioral verification sidesteps the whole problem. It doesn't matter how human-like AI-generated text becomes, because the verification isn't about the text. It's about the observable, measurable process of creating it. You either typed an essay keystroke by keystroke over two hours, or you didn't.
We think schools need to stop tolerating a system that routinely harms innocent students as collateral damage. There's a better way, and the research supports it. Limudi proves student authorship through keystroke dynamics, revision tracking, and cadence analysis, with no content-based guessing and no bias against ESL writers.