Substack Just Gave Your Readers a Button That Scans You

Substack now lets any reader scan a post, note, reply, or comment and see an estimate of how much of it was written by hand versus AI. Substack CEO Chris Best announced it in a post called “Against Claudefishing” — his term for content that fakes a human being on the other end when there isn’t one. Alongside the scan, every writer can add a “How I make this” statement: an optional note, attached to your publication, that explains your actual process. Write that statement. It’s the one part of this whole rollout you get to control completely.

I found the setting completely by accident. I was building a Substack support tool, poking around in a client’s account settings, and there it was — a field called “How I make this.” What the what?

Not gonna lie, my first reaction was not the confident one you’d want from someone who’s taught this stuff — marketing, branding, writing — for over twenty years. I write with AI every single day. I have a clone trained on my own voice. I’ve taught more women than I can count how to do exactly what I do. And my very first thought, before any of that, was to wonder what the scan would say about last week’s newsletter. Crapadoodle.

So I did the thing I tell everyone else to do, and actually read the post instead of just panicking about it.

What did Substack actually launch, and why?

Substack launched a scan, built with a company called Pangram, that runs on any post, note, reply, or comment over a hundred words and shows an estimate of human versus AI to whoever asks for it. Writers get their own side of it too — you can run Pangram on a draft before you publish, add your “How I make this” statement, and report a scan you think got you wrong.

Best’s stated reason is that an increasing share of what people scroll past online was made by no one, and he cites Pangram’s own estimate that it’s as much as 40% on some platforms. (LinkedIn, anyone? To be fair, that’s Pangram’s number about Pangram’s own market, not an independently checked figure, so make of it what you will.) His actual worry, underneath the number, is a mismatch — a reader invests real attention assuming there’s a real person behind the writing, and there isn’t one. That’s the con. Not the AI. The mismatch.

Two more things worth knowing. He says outright these tools don’t train on your writing, which is the first thing I teach the women I coach the second an AI feature shows up on a platform they publish to. And within days of launch, based on writer pushback, Substack changed the feature so you can turn scanning off on a single post without ever touching Pangram — which tells you two things at once: they were listening, and a genuinely large number of people wrote in furious.

Is this actually anti-AI?

No, and I went in expecting it to be. Best writes plainly that Substack isn’t against AI-assisted work, that people should get to choose their own tools. Near the end of the post he writes: “when I want Claude’s opinion, I’ll ask Claude.” Use the tool — but the opinion at the end has to be a person’s.

I think that’s the correct line to draw, and I say that from inside a business that runs almost entirely on AI-assisted work. Nobody has settled etiquette for any of this — there’s no rulebook, no governing body, we’re all making up the norms in public at the same time — so any position a platform takes is going to make somebody furious. Substack picked disclosure over prohibition. That’s the one I’d have picked.

What actually goes in a “How I make this” statement?

It’s a short note, attached to your publication once, that shows up next to your scan result and tells a curious reader how you actually make your work. And the ones people have already published are a lot better than I expected, because there’s no house style and nobody is grading you on it.

The economist Claudia Sahm wrote that her first draft is always her own, that AI mostly helps her with clarity in editing, and that any errors in the piece are hers regardless. A writer publishing as Thierry, from arvy, said English isn’t his first language and he uses AI the way any writer works with a copy editor — shaping the prose, never the substance. A writer named Nate said flatly that if a scan flags his work as AI-assisted, that’s accurate and expected. And then there’s Plum Sykes, whose entire disclosure is “No AI to see here darlings!” — I want to be her when I grow up. (These, and the ones below, are quoted in Substack’s own follow-up post on the launch.)

The post I liked most came from Sara Szal, a physician who wrote that every idea in her newsletter comes from thirty years of clinical practice and six published books, and that she uses AI trained on her own writing to edit and publish more efficiently — the same way she’d use any other tool in her practice — while the judgment, the clinical thinking, and the framework stay entirely hers. Decades of expertise stated first, the tool named plainly second, not one word of apology anywhere in it.

Can the scan just be wrong about you?

Yes, and Best doesn’t dodge it — Pangram isn’t perfect, he says so directly, even while pointing to third-party evals showing it’s accurate most of the time. One woman, Monica Hebert, published her own response calling the whole feature “another purity test.” Her point: a detection score can tell you a tool was involved, but it “cannot tell you who had the original thought, who shaped the argument, who rejected the bad language, who supplied the life, or who stands behind every word.” She’s been publishing for eighteen months, she said, and not one reader has ever accused her of running slop.

So hold both truths at once: the scan will be wrong about somebody this month, possibly you — and that’s the argument for getting your own statement up before anyone runs a number on you.

Why is this a trust problem, not a technology problem?

KLT — know, like, trust — is the whole foundation under any visibility system worth building. What Substack just did was bolt a trust mechanic straight onto the platform. But trust was never supposed to be the thing a piece of software could measure for us.

Your readers already like you. They like your writing, no matter how you create it. So the readers who already know you probably aren’t the ones running that scan. The person actually pressing that button is someone new, three paragraphs into your post, deciding whether you’re worth the rest of their morning. Your statement is doing that work before “like” and “trust” ever get a chance.

How do you write your own statement?

Five sentences, ten minutes, and you’re done. Work through it in this order:

1. Name where the idea came from — you, not the tool.
2. Say specifically where AI comes in. Research, drafting, editing, cleaning up transcripts. Don’t be vague about it.
3. Say what you never hand over. For most of the women I work with it’s the client stories, the judgment calls, the opinion — the part that would be wrong if someone else had written it.
4. Say whether a human reads every word before it publishes. If yes, just say so.
5. Close on your credential. Twenty years. Six books. A practice you built from nothing.

Steal the structure, not the sentences. Yours should sound like you, which is sort of the entire point of the exercise.

Two ways people mess this up

Writing it like a legal disclaimer — four hedged sentences. Just talk like you would over coffee.

Or writing it once and never touching it again. My own process has changed twice this year. Reread yours every January.

AI is smoother than you’ll ever be, creamier and breezier and all of it, and your sharp edges are the entire point — don’t go sanding them off just to sound like the robot, and don’t hide the fact that you use one either, because the hiding was always the only part of this that could cost you anything.

Write your statement this week. Then go tell your freaking stories.