Can AI Detectors Be Wrong? What the Research Actually Shows

Quick answer

Yes, unambiguously. Every published independent study of AI-text detectors — not just older tools, the current generation too — finds real false-positive rates, and they are not evenly distributed. Non-native English writers get flagged at dramatically higher rates than native speakers writing the same kind of content. That's not a reason to throw detection out. It's a reason to use a score as one signal in a human decision, never as the decision itself. This piece walks through what the research actually measured, what happened when schools ignored it in 2025, and what a detector score should and shouldn't be used for.

The core finding: detectors don't fail randomly, they fail predictably

The most-cited study on this, from a team of Stanford researchers led by James Zou, tested seven commercial and open-source AI detectors on two sets of essays: writing from U.S.-born eighth graders, and TOEFL essays written by non-native English speakers (Stanford HAI, "AI-Detectors Biased Against Non-Native English Writers"; underlying paper: Liang et al., "GPT Detectors Are Biased Against Non-Native English Writers," arXiv).

The detectors were "near-perfect" on the native-English eighth-grade essays. On the TOEFL essays — all genuinely human-written — they misclassified 61.22% as AI-generated. Worse: all seven detectors unanimously flagged 18 of the 91 TOEFL essays (19%), and 89 of the 91 (97%) were flagged by at least one detector.

Why the gap? Detection leans heavily on perplexity — a statistical measure of how "predictable" word choices are. Sophisticated, varied vocabulary and complex sentence structure score as less predictable (more human-like) under this metric; simpler, more formulaic phrasing scores as more predictable (more AI-like). Non-native writers, on average, produce lower lexical and syntactic complexity than native speakers not because a machine wrote their essay, but because they're writing in a second language. The detector isn't measuring "was this written by AI." It's measuring "does this resemble the kind of text AI tends to produce" — and formulaic, safe, simple writing resembles that whether a human or a model wrote it.

Zou's team also demonstrated the flip side of the same weakness: genuinely AI-generated text can be rewritten with a single prompt — asking a model to "elevate the provided text by employing literary language" — and slip past the same detectors that just flagged an honest human essay. A system that's simultaneously too suspicious of simple human writing and too easily fooled by dressed-up AI writing has a real design problem, not just an edge-case bug.

This isn't a 2023 problem that's been fixed

It would be reassuring to say the 2023 study describes an early generation of tools that's since been corrected. The evidence says otherwise. Reporting from December 2025 found the same pattern still playing out in real classrooms, at scale (NPR, "Teachers are using software to see if students used AI. What happens when it's wrong?").

More than 40% of surveyed 6th- to 12th-grade teachers used an AI detection tool during the 2024-2025 school year, per a nationally representative poll by the Center for Democracy and Technology. Mike Perkins, a researcher on academic integrity at British University Vietnam, told NPR his testing found popular detectors — he named Turnitin, GPTZero, and Copyleaks specifically as tools he'd evaluated — both flagged human writing as AI and missed AI writing that had been lightly edited to look more human. "It's now fairly well established in the academic integrity field that these tools are not fit for purpose [as a sole basis for accusations]," Perkins said.

Turnitin's own public position, quoted in the same NPR piece, is that "our AI writing detection may not always be accurate … so it should not be used as the sole basis for adverse actions against a student," and that it's more important to avoid falsely accusing an honest student than to catch every instance of AI writing. That's a notable admission directly from a detection vendor, not a critic.

What false positives look like in practice

Two documented 2025 cases show the human cost when a raw percentage is treated as a verdict instead of a prompt for a conversation:

• A University at Buffalo student, Kelsey Auman, was accused of academic dishonesty on multiple assignments in April 2025 after Turnitin's AI detector flagged formulaic, citation-heavy writing — a genre that reads as "predictable" almost by definition. She found roughly 20% of her class had also been flagged. Her case was eventually cleared after she produced her browser history and drafting research, days before a graduation deadline (Spectrum News 1, "Western N.Y. student's AI use accusation questions validity").
• A 17-year-old Maryland high schooler, Ailsa Ostovitz, was flagged in September 2025 on an assignment where she wrote about music she personally listens to — a 30.76% AI-probability score from a detector her teacher was using independently, outside any district-purchased tool. The district later confirmed it doesn't pay for or endorse that software and advises staff not to rely on such tools, per NPR's reporting.

Both cases resolved in the student's favor. Neither resolved quickly, and neither would have needed to happen if the score had been treated as a starting point for a conversation rather than a finding.

Separate reporting on the Auman case notes at least three universities have already disabled or downgraded reliance on AI-detection scoring over exactly this bias: Vanderbilt disabled Turnitin's AI detector outright in 2023, citing false positives and elevated flagging of non-native English writers; the University of Nebraska-Lincoln has flagged higher false-positive rates among neurodivergent students; and MIT's own teaching-and-learning guidance argues AI detectors "don't work" reliably enough to anchor a policy on (Spectrum News 1).

So what is a detection score actually good for?

None of this means detection is useless — it means a single number from a single tool is the wrong unit of decision-making. Used correctly, a detection score is:

• A prioritization signal, not a verdict — a high score on one submission out of two hundred tells a reviewer where to look first, not who's guilty.
• One input among several — writing-process evidence (drafts, version history, browser activity, prior writing samples from the same author) resolves ambiguous cases far more reliably than a percentage alone, exactly as it did in both cases above.
• Most useful at the extremes, least useful in the middle — a 2% score and a 98% score both carry real information; a 30-45% score sits in territory where most published research says the tool itself is least confident, and treating it as a near-accusation is where harm concentrates.
• A conversation starter, never a grade — asking "can you walk me through how you wrote this" costs nothing and catches the false positives that a hard cutoff can't.

This is also why TheChecker.AI reports 93% accuracy across 40+ source models as a measured, published figure rather than an implied guarantee of certainty on every single submission — an aggregate accuracy rate describes how the tool performs across a large test set, not how confident you should be about any one specific score sitting in ambiguous territory. Good detection tooling should make that distinction easy to see, not bury it in a single percentage. If you want the mechanics behind why any detector — ours included — leans on signals like perplexity and burstiness in the first place, how AI detectors actually work breaks that down in plain terms.

Try it on your own writing before you rely on someone else's

The fastest way to understand what a detection score does and doesn't tell you is to run a piece of your own, verifiably human-written text through a detector and see how it scores — then try a heavily-edited AI draft and see how that scores too. TheChecker.AI's free demo lets you paste up to 1,000 characters and see the sentence-level breakdown behind the topline number, so you can see where confidence is high versus where it's genuinely uncertain, instead of getting a single opaque percentage. If you're evaluating detection tools for a classroom or a team, that transparency is exactly the thing worth testing for before you trust a score enough to act on it.

FAQ

Are all AI detectors equally unreliable? No — accuracy varies by tool and by how the detector is built, and vendors publish different tested accuracy rates. But the pattern of bias against non-native English writing and formulaic-but-human text is documented across multiple independent detectors in the Stanford study, not isolated to one product.

Should schools stop using AI detectors entirely? Some already have — Vanderbilt disabled Turnitin's AI detection in 2023 for this reason. Others, like Broward County Public Schools (per NPR's reporting), keep using a detector but explicitly as a conversation-starter for teacher-student feedback, not as an automatic grading or disciplinary action.

What should I do if I'm accused based on a detection score? Both documented cases above were resolved with process evidence: draft history, version timestamps, browser activity, or research notes. Keep that evidence as you write, not after you're asked for it.