What AI Detectors Cannot Reliably Determine
AI writing detectors have high false-positive rates, disproportionately flag non-native English speakers, and lack the statistical reliability their users assume.
A jury in California recently found Instagram and YouTube liable in what has been called a landmark social media addiction trial. That verdict drew headlines about platform accountability. But another form of algorithmic judgment is operating with far less scrutiny, fewer safeguards, and more direct consequences for the people it evaluates.
AI writing detectors claim to distinguish human text from machine-generated text. They are deployed by universities, publishers, employers, and social media platforms. And they are demonstrably unreliable.
What detectors actually measure
The tools most institutions rely on — Turnitin’s AI Writing Report, GPTZero, Pangram, and CopyLeaks — share a common analytical foundation. They evaluate text using metrics called “perplexity” and “burstiness.”
Perplexity measures how predictable a passage is, given the statistical patterns the model was trained on. Lower perplexity means the text follows common patterns more closely. Burstiness measures variation in sentence length and complexity. The underlying assumption is that AI-generated text tends to be more uniform and statistically predictable than human writing.
The University of California, Los Angeles has published an analysis of these metrics, noting that they capture surface-level statistical properties rather than any reliable signal of authorship. UCLA’s analysis describes how detectors look for repetitive terms, shifts between formal and informal tone, nonsensical phrasing, and low “unpredictability” — none of which are exclusive to AI-generated text.
The problem is that these properties are not exclusive to AI-generated text. They are also properties of certain kinds of human writing.
The bias problem
A 2023 study from Stanford’s Institute for Human-Centered Artificial Intelligence tested seven AI detectors on essays written by U.S.-born eighth-graders and by non-native English speakers taking the TOEFL exam. The results were stark.
The detectors performed near-perfectly on the native-speaker essays, correctly identifying them as human-written. But they flagged 61.22% of the TOEFL essays as AI-generated. Nineteen percent of those TOEFL essays were unanimously classified as AI by all seven tools tested. Ninety-seven percent were flagged by at least one detector.
The Stanford study attributes this bias to the reliance on perplexity, which penalizes lower lexical and syntactic complexity — features common in non-native writing but not exclusive to machine-generated text.
The University of Nebraska–Lincoln has documented that neurodivergent students — those with autism, ADHD, dyslexia, and related conditions — are also disproportionately flagged by AI detectors. UNL’s research note describes the same mechanism: atypical sentence structure and vocabulary patterns trigger false positives.
The vendors acknowledge the problem in their own documentation. Turnitin published a blog post titled “New Research: Turnitin’s AI Detector Shows No Statistically Significant Bias Against English Language Learners,” but the study it cites uses its own proprietary model and methodology. Turnitin’s claim has not been independently replicated. CopyLeaks similarly markets itself as “The Most Accurate for Non-Native English Speakers” — a claim that addresses the problem while implicitly confirming it exists. CopyLeaks’ marketing does not constitute evidence of resolution.
The human cost
The statistical failures translate into concrete harm.
Jerry Falade, a British-Nigerian author, had a $2 million book deal for a crime novel collapse after his agency withdrew it. The agency stated it could no longer authenticate “how the manuscript evolved.” No concrete evidence of AI use was presented. The doubts arose from an editor’s concerns and changed story details during development — a normal part of the revision process. Falade denies using AI and has pointed to what he describes as racial bias in how his work was scrutinized. The Guardian covered the collapse, and The Wall Street Journal reported Falade’s denial.
Thierry Rignol, a French national studying at Yale University, was suspended after a professor used GPTZero to accuse him of using AI on an assignment. Rignol filed a 13-count federal lawsuit against Yale, alleging that the tool unfairly targets non-native English speakers. Ars Technica reported on the lawsuit.
Orion Newby, a student at Adelphi University, sued after Turnitin flagged his paper as AI-written. A federal judge ruled the plagiarism finding “without merit” and ordered the charge removed from his record. Inside Higher Ed covered the ruling.
Kat Tenbarge, a journalist, was accused by Jack Osbourne of using AI in a Rolling Stone article based on results from Getsolve, an AI detection tool. Tenbarge publicly refuted the claim with evidence of her drafting process. Osbourne did not retract the accusation. Tenbarge’s refutation laid out her notes, outlines, and revision history.
These are not isolated incidents. The New York Times and BBC have published investigations into students falsely accused of AI use. The New York Times and the BBC documented cases where detector results led to suspensions, failed courses, and damaged reputations.
What institutions are doing
The response from universities has been uneven but trending toward skepticism.
Yale University, Johns Hopkins University, Vanderbilt University, and Georgetown University have disabled or restricted AI detection tools on campus. MIT Sloan School of Management has published a direct warning that “AI detectors don’t work,” recommending instead clear policies, engaging assignment design, and inclusive assessment methods. MIT’s guidance does not call for eliminating academic integrity standards — only replacing unreliable tools with more effective ones.
The University of Chicago suggests slowing down reading, breaking up assignments into smaller pieces, and requiring reflection as alternatives to detection. UChicago’s strategies focus on making the writing process visible rather than trying to classify its output. Stanford recommends in-class assessments. MIT advises allowing students to disclose AI use without penalty.
These approaches share a common insight: the problem is not that students are using AI. The problem is that institutions lack a reliable way to distinguish authorized from unauthorized use, and the tools they have adopted create more problems than they solve.
Why deployment outpaced evidence
The Center for Democracy and Technology surveyed U.S. teachers and found that 43% of sixth-through-twelfth-grade educators regularly used AI detectors between 2024 and 2025. CDT’s survey documents widespread adoption despite the known limitations.
The vendors have incentives to project confidence. Turnitin claims a false positive rate of less than 1% for human-written content. Turnitin’s stated rate is measured against its own test sets under controlled conditions. Pangram claims one false positive in 10,000 evaluations. Pangram’s stated rate similarly relies on internal testing. GPTZero publishes disclaimers acknowledging that results should not be used as sole evidence of misconduct, while simultaneously marketing itself to institutions that do exactly that. GPTZero’s disclaimer is buried in its terms rather than presented at the point of use.
OpenAI attempted to build an AI detection classifier and shut it down in 2023 after finding it had low accuracy and would disproportionately flag non-native English speakers. The Verge covered OpenAI’s shutdown, noting the company concluded the tool was “not accurate enough” and could cause harm.
The broader pattern
The deployment of AI detectors follows a recognizable pattern: a tool is marketed as solving an urgent problem, institutions adopt it under pressure, and the limitations become clear only after real people have been harmed by its errors.
Social media platforms have extended the same logic to public discourse. LinkedIn added a “seems like AI slop” button in July 2026, allowing users to flag posts they suspect are machine-generated. The Verge covered the feature. The feedback is used to refine algorithms — a crowdsourced proxy for detection that carries the same fundamental problem: human suspicion is not evidence of machine authorship.
The publishing industry has responded with certification. The Authors Guild launched a “Human Authored” badge to distinguish human-written books, creating a public database of verified works. The Verge reported on the initiative. Independent badges like “Not by AI” and “Written by Human” have emerged as parallel efforts. notbyai.fyi hosts one such registry. These are transparency mechanisms, not detection tools — an implicit acknowledgment that verification requires process evidence, not statistical classification.
Wikipedia created a guide to identifying signs of AI writing and banned AI-generated articles. The Verge covered Wikipedia’s ban. The guide describes patterns like inflated importance claims and superficial analysis — heuristics that overlap with what makes poorly written human text look suspicious.
What remains uncertain
Several questions do not have settled answers.
It is unclear whether the underlying statistical approach can be improved to the point of reliability, or whether the task itself — distinguishing human from machine text based on perplexity and burstiness — has a fundamental ceiling. The metrics capture properties that both humans and machines can produce in overlapping distributions.
It is also unclear how detector accuracy changes as AI models improve. If newer models generate text with higher perplexity and more varied sentence structure, the gap between human and machine text narrows further. The detection problem may be inherently moving target.
The legal standards for using detector results as evidence of misconduct are still developing. The Adelphi ruling established that a Turnitin flag alone is insufficient. The Yale lawsuit may clarify whether deploying a known-imperfect tool against a non-native speaker constitutes discrimination. These cases will take time to resolve.
What the evidence supports
What is clear is that AI detectors, as currently deployed, produce false positives at rates that matter. They disproportionately flag non-native English speakers and neurodivergent writers. They have cost people book deals, academic standing, and professional reputations. The institutions that have examined them most carefully — MIT, Yale, Stanford, Georgetown — have concluded they should not be used as sole evidence of misconduct.
The vendors acknowledge limitations in their documentation while marketing confidence to their customers. The tools are being used for decisions their accuracy does not support. And the people most likely to be flagged are not the people who generated the text with a machine.
That is not a prediction or an inference. It is what the published research, the court rulings, and the institutional responses say.