I make an AI writing tool for a living, so read what follows with one eyebrow up. Here is the claim, which you can test in about ten minutes on your own work: the tools that say they can tell whether a person wrote something are guessing, and when they guess wrong, they do it in a pattern.
What set me off was a prospective customer who told me our sample content "failed AI detection," as if that settled the matter, and it made me ask what those tools actually measure and whether the referee everyone cites can read.
What the tools are measuring
A detector does not sense a human presence. It scores statistical smoothness, so text that is predictable, evenly paced and clean reads as machine, while knotty, uneven, lexically busy text reads as human. That is why a long, cadenced, comma-heavy sentence from the eighteenth century sails through and a two-line business email does not. The email was short and clear, which is what a good email should be.
You may remember the screenshot from 2023 in which a detector flagged the U.S. Constitution as AI-written. The explanation was fair. That document appears so often in the data language models learn from that the models learned to produce text just like it, so a tool built to spot model-like text trips on the original. A tool patched to clear the exact documents that once embarrassed it, while still flagging a plain email, has studied the answer key, and the method underneath still cannot read.
It is worse for people who learned English second
Anecdotes prove little, so here is peer-reviewed evidence. In 2023 a Stanford group led by Weixin Liang and James Zou tested seven detectors against essays written by real people, and on writing by native English speakers the detectors were close to perfect. On TOEFL essays written by non-native speakers, they falsely flagged a large share as AI-generated, and at least one flagged almost every essay in the set. The study ran in the journal Patterns, and the finding is blunt: these tools are biased against people who did not grow up speaking English.
The mechanism is uglier than a bug. Essays that got flagged used simpler words and steadier sentence structure, and the detectors read that plainness as a machine. Students most likely to be falsely accused of cheating are the ones still building their vocabulary, and the fastest way for them to look human to the software is to run honest work through a chatbot to fancy it up. A tool that claims to catch AI use ends up arguing for more of it.
What a machine cannot do
Malcolm Gladwell built a whole book, Blink, around a moment at the Getty museum. Experts looked at a marble statue the museum had authenticated through fourteen months of testing, and in the first two seconds something in them said wrong. They could not always explain it on the spot, and they turned out to be right: the statue was a fake. Gladwell's point was that a trained eye compresses years of pattern into an instant of judgment, and that this judgment often beats the slow machinery built to replace it.
A good editor does the same with your draft. She knows, fast and by feel, whether a person with something to say is behind it. A detector does arithmetic on sentence shapes instead, then hands you a percentage precise enough to sound like certainty. Precision and accuracy are different things. A scale that always reads 12.4 pounds is precise, and still broken if the bag weighs five.
So we do not sell you a number
A marketer is supposed to say here that the writing beats the detectors, and I will not, because chasing a score makes writing worse. Plain, direct prose often fails these tools while padded, over-complicated prose passes, and anyone who promises a particular result is promising something that no one, least of all a vendor, can control.
At Ghosts the work goes into the things a reader notices. Your writer is trained on your own samples, so the voice is yours. The humanizing passes are the ones a working editor would make: they vary the rhythm, swap stock phrases for plain words, cut hedging, and take out claims with nothing behind them. Every fact is checked against the sources the research found. The test we believe in is older and simpler. Read the draft, and ask whether you would put your name on it and send it.
Do not take my word for any of this, since I have a tool to sell. Run your own writing through any detector you like, watch what it says about the most human thing you wrote this week, and decide for yourself who the referee is working for.
FAQ
Why do AI detectors flag writing that a human clearly wrote?
Most detectors score statistical smoothness, not authorship. Clear, evenly paced,
plainly worded text reads as machine-generated to them, which is why short human
emails and essays by non-native English speakers get falsely flagged.
Are AI detectors accurate?
They stay unreliable. A 2023 Stanford study found detectors falsely flagged a large
share of essays by non-native English writers as AI, while clearing native-speaker
writing almost perfectly.
Can I trust a detector score to prove my work is human?
No. The score reflects sentence complexity more than authorship, so a real person
can fail while edited machine text passes. Editorial judgment, meaning whether
a person with a point of view is behind the words, is the more honest test.