The morning began like any other. I finished a piece I had turned over in my head for days, edited it until the lines said what I wanted, then hit publish. Every idea in it was mine, from the first thought to the final word.
Then the platform handed down a verdict. A detection tool had scanned the piece:
AI 100%. AI-assisted 0%. Human 0%.
A tool that had never read a word I’d written decided, in a fraction of a second, that none of it was mine.
Let me be direct about something the software couldn’t know. I do use AI. I wrote a book about it and speak and teach about using it well.1 That is exactly why this is worth noting. The tool didn’t catch me at anything; it made an assumption and labeled it as proof.
The accepted version
It goes like this. Detectors protect honest writing. A flag means someone cut corners, handed a machine their thinking, and passed it off as their own. You see a percentage, trust the number, and move on.
It is a tidy explanation—and it is also selling something.
Under the number lies a question nobody asks: what would it take to prove a piece of writing is human? The answer lives below the surface. The idea nobody handed you. The judgment about what to cut. The hours of editing. The voice that took years to find. That is what makes writing yours.
A detector cannot see a single one of those things.
What the machine really does
Here is what it can see. Patterns. It reads the shape of your sentences and compares them to what it was trained to call “machine.” It does not know if you meant it. It does not know if you bled over it. It knows whether your writing is consistent.
That’s the trap. Good writing has habits. Clean prose. Tight structure. Even the same em dash, used the same way every time. I’ve been doing that for years. Ironically, the more disciplined your writing becomes, the more likely a machine is to accuse you of being one.
So the writer who edits hardest gets flagged fastest. The reward for polishing your voice is a decision that says you never had one.
Patterns aren’t provenance.
Sold as proof
Then there is what the detector actually is. Not a referee. A product. It is software sold to platforms and readers, and it sells certainty: a number you can point to and say, “See? Proof.”
But a determination only means something if it is right. And these tools are wrong often enough that the people who build AI do not trust them.
In early 2023, OpenAI—the company behind ChatGPT—released a tool to detect AI-generated writing.2 A few months later, it quietly shut it down, admitting it was too inaccurate to stand behind. The company that makes the machine could not build a reliable tool to catch the machine. It gave up.
That is the admission. The maker of the AI did not believe in the detector. Yet here we are, treating a lesser one as gospel.
The AI nobody complains about
Step back and look at how much AI you already trust without a second thought.
Autocorrect finishes your sentences before you do. Grammar tools rewrite your lines, and you take the suggestion. Your GPS reroutes you around traffic it predicted before you left the driveway. Spam filters decide, silently, which mail reaches you at all. Your devices read articles aloud in a synthetic voice no one ever recorded, and they have for years. Schools have run student papers through plagiarism software for more than a decade—pattern-matching tools, the very same family as the detector now failing me. And a growing share of your searches end with an AI summary sitting on top of the results, answering before you click a thing.
None of it wears a label. Nobody demanded one.
Then the written word alone got the scarlet letter.
Here is the contradiction at the center of it. The thing judging your writing is itself AI—software trained on human words, deciding which human words look too much like a machine. It flags your sentence as fake. Then a reader arrives through an AI summary, on a platform that will gladly read your post aloud in an AI voice. AI accuses. AI narrates. AI answers. You are the only one in the whole chain made to prove you are human.
What this trains you to do
This is where it stops being about one flagged essay.
A reader who lets a detector decide what is human has handed over their own judgment. That is the same surrender I traced in The Auction Was Over Before You Opened Your Phone—the answer engines that hand you a reply before you can weigh anything yourself.3 Stop reading for yourself, stop weighing the source, and let the number tell you what to believe. That is the whole game.
And the writer learns to flinch. You start bending your sentences to evade the software’s suspicion. You rough up your best lines on purpose so they read as “human enough.” You write worse to please a tool that cannot read.
Convenience becomes a habit. Habit becomes conditioning. One more decision handed off, this time about what is real.
Test it yourself
So what do you do about it? You test. You weigh it yourself. You do not take the label’s word for it. The practice is not complicated.
Disable the scan. Substack lets you turn it off, piece by piece, and replace it with your own words. Trade the machine’s call for your standard.
Explain how you work. Write a short note that tells readers the truth: the ideas are yours, the effort is real, and the tools serve the thinking, never replace it. Own your process out loud, and the accusation loses its teeth.
Keep the record. Draft where your revision history is saved. A timestamped trail of your edits is evidence in a way a guess never will be.
And the big one. Decide for yourself. Do not let a detector tell you or your readers what is true. Read it. Weigh it. Judge it. Your ear is the finest detector ever made: it catches the phrase no person would say, the line that rings false—the tell the machine keeps missing. That authority belongs to you.
A detector can flag a pattern. It cannot read you, your intent, your labor, or your voice on the page. The moment you let it, you have handed over the one thing no machine can give back: the authority to say what is true.
That authority was never the machine’s to take. It is yours to protect.
Note: The “verdict on a paper it never read” framing and the reading/authorship metaphor are my interpretation, not the detector’s own language. My claim is not that AI detection never works. Rather, these tools cannot measure authorship, effort, or truth, and should not be sold to readers as if they can.