The AI detector on Substack can't find the driver. It measures the road, not who's steering.

I scanned some of the people I follow. It catches the format, noting else.

After reading the article by my business coach and mentor, Suzanne Taylor-King “Substack Installed a Lie Detector,” I have some thoughts.

You’ve probably seen it by now. A section that tells you how the article was written (sort of).

AI-assisted. Human. A percentage next to each one, telling you and everyone else how much of what you just read came from a machine.

It’s called Pangram. Substack partnered with them to scan posts and score the writing.

I went down a rabbit hole with it this week. Started checking people I actually subscribe to. People I respect. And the results were, let’s say, interesting……

Let me tell you what I found. Then let me tell you why almost everyone is reading it wrong.

First, what this thing actually is

The detector does not read your mind. It cannot watch you work. It reads the words on the page and asks one question. Does this look like raw output from a model?

That’s it. Pattern matching. Polish. Structure. Rhythm. Whether the prose has that clean, evenly paced, nothing-out-of-place quality that unedited AI tends to produce.

Then it hands you a score. AI. AI-assisted. Human.

Substack is drowning in the same slop as everywhere else. Thousands of newsletters that are a prompt and a publish button. If a detector gives readers an early smell test on that, good. We want that. The whole reason people ran to Substack was to get away from the algorithmic sludge in the first place.

So the tool is not the villain here. Stay with me.

What you’ll see when you go looking

I scanned a handful of people. Here’s the spread.

Dan Martell. Massive audience. His full guide on building AI agents came back 100% AI. Zero human.

Sabrina Ramonov, the person I’d point a beginner to if they wanted to actually learn this stuff, 98% AI on her SEO playbook. Two percent human. The subtitle promises copy-paste prompts for every step, so, yeah.

Suzanne Taylor King, my own coach, 100% AI on a recent one. And she knew. She wrote a whole piece about getting flagged, titled “Substack Installed a Lie Detector, and It Is Pointed Directly at Me.” She didn’t scrub it and go quiet. She led with it.

Then a good friend of mine, Gaia Girus-Nowoczyn. 67% human. Mostly human-written. She’d pinned a note explaining exactly how she works. She takes what the AI gives her and argues with it until there is nothing left to push against. Her words, not mine.

And Ruben Hassid? Couldn’t even scan him. His post sat behind the paywall, and the detector only reads the free preview. So no score. Just a lock.

Oh, and video? Not checked. The detector reads text. Everything you say on camera is still the wild west.

So already you can see it. Some people get a number. Some you can’t check at all. Some are honest about the number before you even find it.

But look at who I was checking

Full disclosure on my own little experiment. My feed is stacked with AI people. Claude, agents, automation, the whole crowd. I subscribe to who I subscribe to, and I lean AI-heavy because that is my world.

So of course my sample skewed high. I went fishing in the one pond most likely to score AI.

Another good friend, JT Your Dog Knows , obviously not AI specific but dog specific, couldn’t use the AI detection. I don’t know why.

Here’s why that matters, and it is not the reason you’d guess.

It is not only that AI people use more AI, although sure, some do. It is the format. Look at how this crowd writes. Numbered lists. Full guides. Full playbooks. 27 tips after 1,800 hours. How-to-build-your-first-anything walkthroughs. That is the single most machine-shaped way to write. Tidy list, even cadence, every point wrapped in a bow. Very English university coded.

Go back to Sabrina’s title. Full Playbook. Of course it hit 98%. A playbook is built to read like a manual. That is the entire job of a playbook. It has nothing to do with whether she cheated and everything to do with the shape of the thing.

The detector does not flag that because it is fake. It flags it because it is built like a manual. And manuals read like machines no matter who typed them.

Which is the whole point.

The detector cannot find the driver

Read those results again.

Dan and Suzanne got flagged 100%. Gaia got flagged 67% human. And the lazy conclusion is obvious. Dan and Suzanne used more AI. Gaia used less.

No.

The detector is not measuring how much AI touched the piece. It literally cannot know that. It’s measuring how much of a human survived in the writing.

Dan’s guide is smooth. Structurally perfect. Not a fingerprint on it. That’s what 100% looks like. It does not mean he prompted and posted. It means whatever he did, none of him made it onto the page.

Gaia landed human for the opposite reason. She fought the draft. The friction shows up in the words. The uneven rhythm. The turns you wouldn’t expect. That is not a lower AI dose. That is a person who stayed in the seat.

The machine sees the road. It cannot see who is steering.

Which is exactly why the shame game is backwards

Here’s what’s already happening. People are screenshotting the 100% scores and waving them around like a gotcha. Caught you. Fake. Fraud.

Think about who that actually punishes.

It punishes polish, not dishonesty. And worse, it punishes the honest ones. Because the person who lies about using AI, who runs it through and then manually roughs it up to beat the score, walks clean. The person who used AI openly and didn’t perform enough mess gets flagged.

The detector rewards the fakers who learned to game it and shames the people telling the truth.

That’s the whole problem in one sentence.

Look at Suzanne again. Flagged 100%, and instead of hiding it she named it and wrote the piece. That is the opposite of the shame game. That is someone refusing to be embarrassed by a broken scale.

So here’s my position. Flag me.

I use AI. I have never pretended I don’t. I use it every day and I am still the one driving.

The ideas are mine. The experience is mine. Twenty-four years of it. The voice is mine. AI is how I move faster. It is not where I outsource my thinking.

So flag me. 100%, 60%, whatever the bar reads that day. It changes nothing about who was in the seat.

Because the detector measures whether I sound like a machine. It does not measure whether a machine made me. Two completely different questions. Everyone is arguing about the wrong one.

I’m not scared of the score.

I’m scared of the day I read something of mine and can’t feel myself in it anymore.

That’s the only detector that counts.

---

If you went score-hunting this week too, whose did you scan? And did the number match what you already believed about them, or did it flip something?

If this is the kind of thing you want landing in your inbox, subscribe. I write about staying yourself while the whole internet automates itself into sameness.