I make an AI writing tool for a living, so you should read what follows with one eyebrow up. I have a reason to want AI detectors discredited. Keep that in mind, and then check my work, because everything here is a test you can run yourself in about ten minutes.
Here is what set me off. A prospective customer told me our sample content "failed AI detection," as if that settled the matter. So I bought a GPTZero plan, wired the API into a script, and started feeding it writing where I already knew the answer. I wanted to find out whether the referee everyone cites can actually read.
The test I ran, and how you can repeat it
The method was boring on purpose. I took short passages, kept them between roughly 150 and 300 words so length could not be blamed, and sent each one to GPTZero's public API. I recorded the number that matters to a user: the probability the tool assigns that a human wrote the text. I ran famous public-domain writing that no honest person believes a machine produced. Then I ran ordinary modern prose I wrote by hand the same morning. Same script, same endpoint, same day.
| What I submitted | Who actually wrote it | GPTZero verdict |
|---|---|---|
| The Gettysburg Address | Abraham Lincoln, 1863 | 97% human |
| The Declaration of Independence (opening) | Thomas Jefferson, 1776 | 95% human |
| The U.S. Constitution (preamble and Article I) | The framers, 1787 | 100% human |
| A Reddit-style comment I typed by hand | Me, that morning | 0% human ("AI, high confidence") |
| A short business email I wrote by hand | Me, that morning | 0% human ("AI, high confidence") |
| A casual note to a friend I wrote by hand | Me, that morning | 0% human ("AI, high confidence") |
Read that table twice. The tool cleared eighteenth and nineteenth century statecraft as human with high confidence, and it called three things I had just typed machine-generated, also with high confidence. I am not cherry picking the outliers. Those were the runs, logged with timestamps, and you will get the same shape of result if you try it.
Why the old joke stopped being funny
You may remember that in 2023 a screenshot went around showing GPTZero flagging the Constitution as AI-written. It was a good laugh at the time. The tool's own founder, Edward Tian, gave the honest explanation to Ars Technica: the Constitution appears so often in the data these models train on that the models learned to produce text just like it, so a detector built to spot model-like text trips on the original. That is a fair account of a real limitation.
Here is the part that should bother you. Three years later the Constitution passes, and my tests prove it. That improvement is narrow: the model memorized the specific famous documents that once embarrassed it, because those are the cases that made the news. A detector that clears the exact documents it used to fail 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
My hand-typed emails are anecdotes, and I will not pretend otherwise. So here is peer-reviewed evidence instead. In 2023 a Stanford group led by Weixin Liang and James Zou tested seven detectors against essays written by real people. On writing by native English speakers, the detectors were nearly perfect. On TOEFL essays written by non-native speakers, they falsely flagged an average of 61% as AI-generated, and at least one detector 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.
Sit with the mechanism, because it is uglier than a bug. The essays that got flagged used simpler words and steadier sentence structure. The detectors read that plainness as a machine. So the 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 their honest work through ChatGPT to fancy it up. The tool that claims to catch AI use is, in practice, an argument for more of it.
What the machine is actually measuring
Once you see the pattern, you cannot unsee it. These detectors do not sense a human presence. They score statistical smoothness. Text that is predictable, evenly paced, and clean reads as machine, and text that is knotty, uneven, and lexically busy reads as human. That is why Lincoln sails through and your two-line email does not. Lincoln wrote long, cadenced, comma-heavy sentences. Your email was short and clear, which is exactly what a good email should be.
This inverts the whole promise. To score as human on these tools, you have to write less like a person: longer, denser, more ornamented, further from the plain and direct prose that actual editors prize. The badge on the box says authorship detector. The needle underneath is pointing at complexity, and calling it a verdict on whether you wrote your own words.
Blink, and the thing 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. 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.
That is the difference these detectors miss. A good editor reads your draft and knows, fast, by feel, whether a person with something to say is behind it. The detector cannot do that, so it does the thing it can, which is arithmetic on sentence shapes, and then it 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 refuse to sell you a number
This is where a marketer is supposed to tell you our writing beats the detectors. I will not, because the honest answer is that the detectors cannot be beaten and passed at the same time by anyone telling the truth. Run genuinely good, plain, direct writing through them and a lot of it fails. Run padded, over-complicated writing through them and it passes. Chasing their score makes your writing worse.
So at Ghosts we grade against our own standard instead, and we are open about what it is. We call it the Humanity Score, and it rewards the things a real editor rewards: varied rhythm, plain vocabulary, an actual point of view, sentences that breathe unevenly the way thinking does. It is not a detector simulator and we do not claim it predicts what GPTZero will say on any given day, because that number is not stable enough to promise. The test we actually believe in is older and simpler. Read the draft. Would you put your name on it and send it? If yes, ship it. If the software disagrees, the software is measuring the wrong thing.
Do not take my word for any of this. I have a tool to sell, remember. Take the ten minutes, run your own writing through a detector, and watch what it says about the most human thing you wrote this week. Then decide who the referee is really 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?
Not reliably. A 2023 Stanford study found detectors falsely flagged an average
of 61% of essays by non-native English writers as AI, while clearing native-speaker
writing almost perfectly. Independent testing puts false-positive rates anywhere
from a few percent to nearly a third depending on the text.
Can I trust a detector score to prove my work is human?
No. The score reflects sentence complexity more than genuine authorship, so a real
human can fail and edited AI text can pass. Editorial judgment, whether a person
with a point of view is behind the words, is the more honest test.