Back in 2023, someone ran the United States Constitution through a popular AI detector called ZeroGPT. The verdict came back: mostly written by AI.
The Constitution. From 1787. Whatever James Madison was using to draft it, it wasn't ChatGPT.
That one example tells you almost everything you need to know about AI writing detectors. They feel authoritative. They hand you a confident percentage. And a lot of the time, they are simply wrong.
How the detectors actually "work"
Most of these tools don't detect AI. They detect predictability. They measure how expected each word is given the ones before it, and if your writing comes out smooth, clean, and low on surprises, they call it a machine.
The trouble is that plenty of human writing is smooth, clean, and low on surprises. So the people who get flagged aren't cheaters. They're:
- Non-native English speakers, who often write in careful, simple, correct sentences.
- Neurodivergent writers, who tend to repeat structures and phrases.
- Anyone taught to write plainly and get to the point.
The tool that's meant to catch dishonesty ends up punishing clarity. Students have been accused of cheating over essays they actually wrote. That isn't a rounding error, that's the tool doing real harm. (If you want the full technical version of why this happens, Ars Technica has a good piece called "Why AI writing detectors don't work.")
And they miss the real AI anyway
Here's the other half of the joke. While the detectors are busy flagging the Constitution, they wave genuine AI writing straight through.
According to data from Pangram, one of the stronger tools out there, a fair amount of AI content still slips past, on the order of one miss in seventy. And their numbers suggest roughly 40% of longer LinkedIn posts are now fully AI-generated. Substack, for what it's worth, comes out cleanest at around 10%.
So the scoreboard reads: flags humans, misses machines. For something that calls itself a detector, that is not a great record.
Leave the em-dash alone
Somewhere in the last two years, the humble em-dash became a suspect. Use one and someone in the comments squints and goes, "yeah, AI wrote this."
Think about how backwards that is. The em-dash didn't turn into a red flag because robots invented it. It turned into a red flag because AI leans on it, and AI leans on it because it was trained on good writing, and good writers have always reached for it. It's the mark of someone who knows how to steer a sentence.
We managed to reach a point where a sign of decent writing became evidence of fake writing. My honest guess is that it swings back, and the em-dash gets its reputation returned with interest.
(Full disclosure: I'm keeping them out of this piece anyway, because right now they set off alarms and I'd rather you read the argument than fight the punctuation. Which, come to think of it, kind of proves the whole point.)
The question that actually matters
"Is this AI?" is the wrong question, for a simple reason. It's nearly unanswerable, and even when you guess right it tells you nothing useful.
Picture someone recording a walkthrough video, handing it to a model, and letting the model shape it into a written post. A detector screams "100% AI." But the ideas are theirs. The demonstration is theirs. The judgment about what mattered and what to cut is theirs. Is that not their work? Of course it is. The tool changed. The authorship didn't.
The questions worth asking are older and better. Is it true? Is it useful? Does it have a point of view, or is it filler? Did a person actually think here, or did they hand the thinking off? You can answer all of those by reading. No detector required.
What to do about it
Two things.
If you're a teacher, an editor, or a hiring manager, stop treating a detector score as evidence. It isn't. Read the work. Ask a follow-up question. Judge the substance, not the smoothness.
And if you want your own writing to land as yours, skip the punctuation paranoia and go the other way. Put yourself in it. A real number, a real story, an opinion you'd actually defend in a room. That's the stuff no model can invent and no detector can strip out. It also happens to be the stuff that makes writing worth reading in the first place.