English is not my first language. I think in Arabic, I spend a good part of my day in Hebrew, and I do almost all of my professional writing in English — a language I chose, studied, and rebuilt a career inside. So I took it a little personally when I learned that an entire category of software has quietly decided that people who write the way I do are more likely to be machines.
That category is the AI detector: the tools that schools, journals, and hiring teams now run text through to decide — with a confidence score and no real appeal — whether a human actually wrote the words. I work on GenAI quality, so I wanted to know what these systems are truly measuring. It is not "AI." It is something more uncomfortable, and once you see it, you can't unsee it.
The finding that should have ended the conversation
In 2023, a group of Stanford researchers published a study in the journal Patterns with a title that doesn't leave much to interpretation: "GPT detectors are biased against non-native English writers."
They took seven widely used detectors and fed them two piles of text. One pile: essays by US eighth-graders — native English speakers. The other: TOEFL essays written by non-native speakers. On the native-speaker essays, the detectors were nearly flawless. On the non-native essays, more than half were flagged as "AI-generated."
Same honest human effort. Opposite verdict. The only variable that moved was whether English was the writer's first language.
"AI" and "predictable" are the same number
A detector doesn't know what AI is. It can't. What most of them really compute is something close to perplexity — a measure of how surprising each word is, given the words before it. Text a model finds easy to predict scores as "machine." Text full of unexpected turns scores as "human."
Now ask who writes predictably. A language model, obviously — it is by construction a machine for producing the most probable next word. But also: almost anyone working in a second language, leaning on a smaller, safer, more common vocabulary because the flashy synonym is a risk they can't afford to get wrong. The non-native writer and the language model land on the same spot of the only axis the detector can see.
So the tool isn't detecting authorship. It's detecting distance from a native, idiomatic, slightly unpredictable English — and charging everyone else for the gap. The same logic quietly catches neurodivergent writers, anyone trained to write plainly, anyone whose voice sits outside the comfortable middle of the training data.
The cruelest detail in the study
Here is where it tips from biased into absurd. The same researchers tried a fix: they took the flagged TOEFL essays and asked ChatGPT to improve the word choices — to make them sound more literary. The false alarms dropped sharply. Polishing honest human writing with an actual AI made it read as more human to the detector.
Sit with that. The student who writes their own essay in plain second-language English gets flagged. The student who launders it through ChatGPT to dress it up sails through. The detector does not catch the cheater and clear the honest writer. It does the exact reverse.
Even the model's own maker gave up
You'd hope the companies building these models could at least detect their own output. They can't. OpenAI launched its own AI Text Classifier in January 2023 and quietly shut it down six months later, in July, citing a "low rate of accuracy." The numbers behind that phrase: it correctly caught about 26% of AI-written text — worse than a coin flip — while falsely accusing human writing 9% of the time.
The organization with the most intimate possible knowledge of how GPT writes looked at its own detector and decided it did more harm than good. If OpenAI can't reliably detect GPT, the browser extension grading a freshman essay cannot either. It is producing a number that looks like evidence and behaves like a horoscope.
We got the whole paradigm backwards
The deeper mistake isn't any single bad detector. It's the assumption underneath all of them: that "was this written by AI" is a property you can read off the finished text after the fact, like a fingerprint. It isn't. The words on the page don't carry that history. Perplexity — the proxy we reach for instead — encodes a cultural expectation about what fluent English should feel like, and then we aim it at the people least likely to match that expectation and call the result integrity.
And notice what we've done to the burden of proof. The accused now has to prove a negative — that they didn't use a tool — against a system that won't show its reasoning. For a native, idiomatic writer that's a nuisance. For everyone else it's a tax on writing differently.
If we actually cared about authorship, we would stop trying to detect it at the end and start preserving it from the beginning: provenance over forensics. A record of how a piece of writing came to exist, held by the person who wrote it, is the only version of this that survives contact with reality. Everything else is a confidence score wearing the costume of proof.
The part I can't shake
The smallest detail in that study is the one that stays with me: the safest way to avoid being mistaken for a machine was to let a machine rewrite you. We built a guard that rewards the exact thing it's meant to catch and punishes the exact thing it's meant to protect. That isn't a bug in one product. It's what happens when you measure the wrong thing and then trust the number anyway.
So I'll end with the question I genuinely can't answer — especially for the multilingual people reading this. Have you ever changed your own writing — reached for a fancier word, added a flourish you didn't mean — just to sound less like a robot to a robot? Because if you have, the detector already won. And it still didn't catch anyone.
Sources: Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). "GPT detectors are biased against non-native English writers." Patterns, 4(7). · OpenAI. "New AI classifier for indicating AI-written text" (January 2023; discontinued July 2023, per the update on that page).