What AI Content Detectors Get Wrong About Human Writing
AI detectors have figured out the perfect business model: sell the disease and the cure in the same breath. Catch-22's Milo Minderbinder would be proud.
By R.R. Hicks · Forlorn Hope Publishing
A Catch-22 Lesson for the Age of AI
Let's start with Catch-22's Milo Minderbinder.
Milo runs the bomber group's mess hall. Milo can't let a deal pass by. He finds a way to create trades and profit seemingly out of thin air because he always sees both sides.
Eventually, he built a global syndicate – M&M Enterprises ("What's good for M&M Enterprises is good for the country").
At one point, he corners the Egyptian cotton market only to find that he can't sell "a single thread of it." His attempt to get his men to eat chocolate-covered cotton goes about as well as you'd expect.
He's about to go insolvent when he finds the answer to his cash flow problems: he contracts with the Germans to bomb his own airfield.
Back to 2026.
If you're like me, you've been getting bludgeoned by Originality.ai pop-up ads. They stand out because of their duality – they are aimed at content buyers and content creators – and their all-out fear-mongering.
Content buyers: "Unknowingly posting AI-generated content can destroy your brand!"
Content writers are assaulted by images of them being called into the office to explain themselves.
There's a distinct middle-school principal's office vibe to the thing that's surely designed to fuel angsty writers' angst – "not only are we disappointed in you and calling your parents, but this is also going on your permanent record."
How AI Detectors Actually Work
Here's what Originality.ai and all the other AI detectors do: they scan writing for statistical patterns that their particular model's algorithms have decided look like a machine; sentence rhythm, transition density, and how predictably one word follows the next.
By the way, never forget that these models have been trained on billions of human-generated words.
Anyway, clean, well-edited prose that moves without friction (yeah, I, too, have no idea what this means) is flagged as low perplexity (ditto), which is the technical way of saying it reads like something a machine would write.
The cleaner the writing, the higher the score.
Which means, in practice, the better you write, the more likely you are to get flagged as AI.
But wait, there's more…
…they also skim for what some coder and/or machine has decided are "tells."
The em dash is [still] apparently one of them.
I've written this before, and I'm sure I'll be writing it again soon: I've been using em dashes since I read Tristram Shandy in college. It was written in 1759.
Em dashes aren't an AI invention, obviously, and their use is strictly human – they are used as a "rest," to give a thought a quick breath.
The em dash is the mainstay of the self-appointed LinkedIn AI vice squads, though it has experienced a few reprieves only to be consistently retried and found guilty again.
Right now it's considered a dead giveaway, the unmistakable mark of a machine.
Especially by Originality.
But the other rules change every hour.
What's a sure tell?
More importantly, how is a writer supposed to know?
Rules Invented in Rooms Writers Aren't Invited To
They can't. Consider the sketch below — The Whitest Kids U' Know, "The New Thing."
WKUK — "THE NEW THING" — SCRIPT
Bill: "Hey, sorry I didn't hang out with you guys last night."
Kyle: "That's alright, so what'd you guys do?"
Darren: Slaps Kyle in the face
Kyle: "What's wrong with you?"
Darren: "Ah, stop it, what was that for?"
Sam: Slaps Kyle in the face
Kyle: "Ah, stop it!"
Sam: "Oh, dude, you weren't there. That's the new thing we all decided on it last night, pretty cool, huh?"
Kyle: "What's the new thing?"
Bill: "That's the new thing." Slaps Kyle in the face
They continue to slap each other as the game spirals out of control because the characters keep accidentally phrasing questions that end with a one-syllable word
Later in the sketch, they begin adding even more absurd, arbitrary rules to their interactions:
Kyle: "Stop it, stop it. I understand this is the new thing, but can't we just hang out like we did before, dude?"
Darren: "Dude, that's another new thing. If you use a conjunction before a preposition, you get hockey punched."
Kyle: "Did I?"
Darren: "Yep. See, the word 'can't' is a conjunction, and the word 'before' is a preposition. You got me."
Kyle: "You got me, you got me. Stop it, stop it. What's wrong with you guys? I can't hang out with you guys anymore without getting slapped and kicked. I don't know, onomatopoeia. What about onomatopoeia? You guys didn't create any rules for onomatopoeia, did you?"
All Three: Instantly slap and kick Kyle simultaneously.
This is how AI 'tells' work. Kyle is us; he missed the meeting and can't avoid the penalties.
The Yoast Test — Or: Thomas Pynchon Passes Everything
In 2018, I wrote a piece about Yoast, the tool people used then to find out whether an algorithm approved of their writing. I hated it (still do). As an admittedly pissed-off lark one night, I ran some great writers through it. Bob Dylan and Hemingway flunked. Dickens was awful. Huck Finn got a yellow light. Stephen King had the worst score, though he was neck and neck with Lewis Carroll – Alice headed down a new rabbit hole.
Who passed? One guy, Thomas Pynchon. Great author and as notoriously difficult to read as he is to find. He got green lights across the board for the chunk of Gravity's Rainbow I fed it. (I purposely avoided the opening line, one of the greatest ever: A screaming comes across the sky.)
Why am I mentioning this now? Because on a whim (or inspiration) I ran that 2018 piece through Originality.ai last week. It came back 88% AI.
"You pay to use a 'tool' that uses AI to make the AI it's feeding you sound less like AI."
The M&M Enterprises of Content Detection
That's when the whole Milo Minderbinder scenario hit me.
Originality.ai (yes, I insist on using its full name because I refuse to grant that there's anything original about this site) is in business to deal with both sides.
Users pay a fee to see if the writing they are getting from students, content writers, colleagues, etc., is AI.
But Originality.ai is not, at bottom, interested in detecting AI writing.
It makes many writers scared shitless that their writing will get flagged and they'll flunk a class, lose a client, be mercilessly heckled at the next Algonquin Roundtable meeting. (Oh, that Dorothy Parker.)
Originality.ai ruthlessly markets a tool that coaches you to write so you don't sound like AI.
Got that?
You pay to use a "tool" that uses AI to make the AI it's feeding you sound less like AI.
Obviously – well, I hope it's obvious – Originality.ai is vested in finding AI writing.
What to Do Instead
Look... just write.
If the idea is solid, the logic solid, connections unexpected, transitions sizzling . . . odds are it might be so tight Originality will flag it as 52% AI.
Just remember, running AI output through a humanizer so it slips past a detector is the opposite of writing.
It adds nothing, knows nothing, means nothing; it just sands off the fingerprints.
What to Do Instead
Look... just write.
If the idea is solid, the logic solid, connections unexpected, transitions sizzling... odds are it might be so tight Originality will flag it as 52% AI.
Just remember, running AI output through a humanizer so it slips past a detector is the opposite of writing.
It adds nothing, knows nothing, means nothing; it just sands off the fingerprints.
Here's the punchline:
No AI writer/detector would ever have come up with Milo Minderbinder as a metaphor (though, of course, bits and parts of Catch-22 have been scraped and re-scraped by AI a million times).
No chatbot, prompted from now until the sun dies, was going to connect a man who dipped cotton in chocolate to feed his mess to a company selling diseases and cures in the same breath.
That connection is mine.
(It helps, of course, that I read Catch-22 at least once every 3-4 years.)
And…the moment this hits our website, it's over.
It'll get scraped, stored, and served back to the next person as a pattern the machine "found."
The one thing these systems cannot do, I just did for them, for free.
Real writers write.
They know they write well, and they know it the only way anyone ever knows it—after the third draft, the fifth, the tenth, the one where you finally cut the paragraph you were proudest of because it was in the way.
If a machine thinks the final product is "clean, well-edited prose that moves without friction, has low perplexity, and is too clean," and has employed the Scarlet Letter of an em dash along the way—screw it.
Not your problem.
The idea speaks for itself, and ideas are outside the purview of AI.
We Write the Pieces AI Can't Find
Creative Legal Writing produces first-person content for law firm partners and solo practitioners — content with ideas, connections, and voice. Not assembled, not optimized for a detector. Written.
We'd Rather Write the Idea Than Beat the Detector
AI detectors will keep inventing new rules. Writers will keep trying to outsmart them. We'll happily stay out of that arms race.
We're interested in something harder to fake: an idea worth remembering.
That's the kind of legal content we write—first-person pieces built on judgment, voice, and connections no detector can measure.
If your firm is ready to publish something that sounds unmistakably human, schedule a conversation.
Frequently Asked Questions
AI content detectors analyze statistical patterns in writing, such as sentence structure, word predictability, and language patterns, to estimate whether text may have been generated by AI. Their results are estimates rather than definitive proof.
Yes. AI detectors can produce false positives, meaning they may identify human-written content as AI-generated. Their accuracy varies depending on the detector, writing style, and the type of content being analyzed.
AI detection scores should be treated as one signal rather than a final judgment. Human review, subject-matter expertise, and the quality of the content remain more reliable indicators of good writing.
Original ideas, personal experience, logical reasoning, and thoughtful storytelling contribute more to authentic writing than simply avoiding certain words, punctuation, or sentence structures.
