Why AI detectors keep flagging human writing (and what to do about it)

In February 2026, a New York judge threw out a university's AI-misconduct finding against a freshman, calling it "without valid basis and devoid of reason." The evidence was an AI detector. And in a peer-reviewed study, detectors flagged 61% of TOEFL essays written by non-native English speakers as machine-generated — every one of them human.

The short version: AI detectors are unreliable, they're hardest on non-native English writers, and the institutions that trusted them are backing away. If your writing has been flagged (or you're worried it will be), the fix isn't to write worse or more defensively. It's to write more like yourself, and to keep the receipts.

I build on-device writing tools for a living, so I think about the "what counts as AI writing" question more than most.

Why detectors get it wrong so often

An AI detector doesn't read your text the way a person does. It looks for statistical fingerprints: how predictable is this text, given a language model? How uniform are the sentences? How common is the vocabulary?

The problem is that "predictable" and "machine-written" aren't the same thing. Clean, grammatical, well-organized writing is predictable — that's why it reads well. A detector sees a tidy essay and can't tell the difference between a careful human and a machine. That's how the US Constitution scores as AI-generated. That's not a joke; it's the actual failure mode.

The bias is worse in one direction. The landmark study here is Liang et al., "GPT detectors are biased against non-native English writers," published in Patterns (Cell Press). The researchers ran TOEFL essays — written by real humans, under exam conditions — through seven detectors. More than 61% of the essays were flagged as AI-written. Nearly all were flagged by at least one detector.

Think about what that means. The detectors weren't catching cheating. They were penalizing people for writing English in a way that's slightly more formal and slightly more predictable, which is exactly what you get when English is your second language and you learned to write it carefully.

The institutions are retreating

Universities are dropping detector tools entirely, moving to other assessment methods, and dealing with lawsuits from accused students. Inside Higher Ed reported on the shift in August 2026.

That tracks with what the research has shown for a while: false-positive rates are high enough that a detector score can't honestly be called "evidence" of anything. Courts are starting to agree — the New York ruling wasn't an outlier, it was a warning shot.

If you're a student or a professional writer, the practical takeaway is this: keep your receipts. Write in a tool that has version history. Keep your drafts. Run flagged text through several detectors, and appeal formally with the published false-positive data. The appeal process works more often than people expect.

The trap: writing "less AI" by sounding less human

Here's where I think most of the advice online gets it backwards.

The standard playbook for "beating detectors" is to make your writing less predictable: swap common words for unusual ones, vary sentence length, add personality. And that advice works, sort of — but notice what it's actually telling you to do: write like a human.

Which raises the question nobody asks: why was your writing scoring as predictable in the first place?

Usually it's one of two things. Either you're writing in a second language, in which case the detector is just wrong and no writing style will fix that, or you've been leaning on AI tools that generate whole paragraphs for you, and the averaging effect has flattened your voice. I wrote about that mechanism in how to write faster with AI without sounding like AI: when you outsource the entire sentence to a model, you get back the statistical average of a million essays, and the average is what detectors (and readers) notice.

The fix for the second case isn't a paraphrasing tool. That's just a second layer of averaging. The fix is to keep writing the sentence yourself and let the tool finish the ones you already started. Your word choices stay yours. The rhythm stays yours. The finished text is still recognizably you, because you wrote the hard parts — the openings, the transitions, the opinions.

What I'd actually do if I got flagged

Don't panic-edit. A detector score is not proof, and institutions increasingly know it. Gather your version history and drafts first.
Run the text through multiple detectors. The variance between tools is huge — the same essay can score 90% human on one and 90% AI on another. That variance is itself useful evidence in an appeal.
Appeal formally, with the published data. The Liang et al. study and the false-positive rates are public. Institutions that still trust detectors are on increasingly thin ice, and saying so politely is more effective than it sounds.
Change the workflow, not the voice. If you were using AI to draft whole paragraphs, switch to tools that finish your sentences instead of writing their own. The output stops reading like the average of the internet, because it isn't.

That last point is where my bias shows: TypeTab is an on-device autocomplete that finishes sentences you've started — it never writes paragraphs from a prompt, so there's no averaging effect to detect. Your writing stays your writing, word for word. But the underlying argument doesn't need my app: any workflow where you keep authorship of the sentence protects your voice better than any paraphraser.

And if you write in English as a second language, the deck is stacked against you, and the data says so. I wrote more about that angle (and how prediction tools can actually help rather than flatten) in autocomplete for ESL writers.

The receipts culture is here to stay

Detectors aren't going to get dramatically better, because the thing they're trying to measure — "was a human thinking this?" — isn't visible in the statistics of the text. Institutions will keep retreating from them. What replaces them is already emerging: process evidence. Drafts, version history, the ability to talk about your own writing.

If you write all day, the move is to make your process legible. Keep drafts. Write in tools that remember what you wrote and when. And keep the sentences yours — not because detectors demand it, but because the alternative is being indistinguishable from the average, which is a worse outcome than any false flag.

FAQ

Are AI writing detectors accurate? No. Independent studies show high false-positive rates, especially for non-native English writers. Over 61% of human-written TOEFL essays were flagged as AI-generated in the Liang et al. study (Patterns, Cell Press). Universities increasingly treat detector scores as unreliable and are moving to other assessment methods.

Why do AI detectors flag human writing? Detectors measure statistical predictability, not authorship. Clean, grammatical, well-organized writing looks "predictable" to them, so careful writers (and non-native English writers in particular) get flagged. Even the US Constitution has been scored as AI-written by these tools.

Can you appeal a false AI-writing accusation? Yes, and it works more often than people expect. Keep drafts and version history, run the text through multiple detectors to show the variance, and cite the published false-positive research in your appeal. Courts have begun reversing academic penalties based on detector evidence, including a 2026 New York case.

Do paraphrasing tools beat AI detectors? Not reliably, and they make the underlying problem worse. Paraphrasers apply a second layer of statistical averaging on top of your text, which flattens your voice further. Keeping authorship of your sentences (using AI only to finish what you've started) protects your voice better than any paraphraser.