Substack Just Added an AI Detector (and Changed Your Contract)
The two Substack buttons that matter now
AI made content dirt cheap. Anyone can crank out a thousand words in ten seconds now.
And you can feel it, right? Every feed, whether it’s on Medium, LinkedIn or X, is the same flat, tidy ChatGPT voice. And people are getting sick of it. Sick of reading something and realizing… nobody was even there.
That’s one of the reasons Substack shipped an AI detector this week and everybody is losing their minds… and honestly?
Some of them should be because some of them have NO idea what’s actually going on.
You’re about to be the one who knows.
Real quick: this is the recap from our Friday Live.
Every Friday, same time, we get on together: 2 PM Berlin. 1 PM London. 8 AM New York. 5 AM LA.
It’s group coaching where I bring ALL the news and you bring the questions. This week? We talked abou AI and it got spicy….
Want in on the next one with me, award-winning marketer and part-time Substack bestseller?
For the next 24 hours it’s 50% off.
Like… 7 bucks a month. Or 70 for the whole year.
You get the weekly Live, every replay, AND the Substack for Beginners course on demand.
After 24 hours? Gone. Back to full price. I’m not being cute, that’s just how long it runs.
There’s a part of this story hiding in the legal fine print and nobody’s talking about it.
What happened in the week of July 20
July 21st. Substack’s CEO drops a post called “Against Claudefishing.”1
Great title. Here’s what it means.
Claudefishing = catfishing, but with a chatbot.
You read something. You think a human made it. You give it your time, your feelings, maybe your MONEY…
…and there was nobody home. Just a machine.
That’s the thing Substack says it’s fighting.
And notice, the CEO did NOT say “AI is evil.”
His actual point: not everything made with AI is slop, and not everything that’s slop was made with AI.
The problem is the mismatch. You expected a human, you got a vending machine.
So they built a button … but they ALSO changed the rulebook the day before.
What the new AI detection button does on Substack
They teamed up with a company called Pangram who makes AI detectors.
Now there’s a button: “Scan for AI text.”
It’s a READER tool. Not a cop. Nobody gets auto-scanned and branded. Somebody has to CHOOSE to click it. And only they see it.
Only stuff published July 21st or later.
Needs 100+ words. Your little 2-line note is safe.
Works on posts, notes, replies, AND comments. Yeah. Comments too.
It’s free.
Web, iOS, Android. Doesn’t work on video, audio, or custom-domain sites.
You get a SPLIT, not a verdict. Human / AI-assisted / AI-generated. Percentages.
The Scan for AI text button on Substack
Wanna scan somebody else’s post?
Three dots at the top right. “Scan for AI text.”
Wanna scan your OWN draft before you hit publish?
Write it , Continue (or Next on mobile), “Scan for AI text”, read your report before the world does, if you care.
The 2 new Substack buttons that matter
There are two settings 90% of people don’t know about.
Button one: “How I Make This”
In our LIVE on Friday, no one had a clue but it’s your public statement about how you work and it pops up when anyone scans you.
Settings_Details_Edit “How I make this” statement and then you write it and hit Save.
WHY it matters:
If you already told people how you work… a “40% AI-assisted” score is just… true. You already said it and there’s no gotcha left.
The score only hurts you if it reveals something you were HIDING or says you’re using AI althoigh you don’t (false-positive results are common and one of the reasons on Reddit there are soooo many folks questioning the tool).
Button two: Disable detection
You can turn the scan off: once you publish a post at the bottom click “Disable AI detection”. Readers just see “AI detection unavailable.”
At launch you had to scan your post first before you could turn it off (annoying, right?). People screamed. Substack fixed it. Now you can disable it without scanning at all.
The catch that’s still there: it’s per-post. No master off-switch yet. That’s the #1 complaint. People want one button for their whole publication to simply opt-out and I think that’s fair. What do you think?
Oh and you can dispute a scan you think is wrong. Keep that in your back pocket. It matters more than you think.
How Pengram works
The old AI detectors? Garbage ?!
They played “guess the next word.” Basically: “if I were ChatGPT, what word comes next?”
Problem: that flags the Declaration of Independence as AI. And Frankenstein. And the Bible and so many more classics.
Why is this?
Computer engineer and AI expert Patrick God, says:
“Because the models read those a million times, so every word looks “predictable,” so the dumb detector goes “AI!””
Pangram doesn’t do that.
It’s a classifier.
“You can picture a giant map of writing styles. Similar writers stand near each other. Human writing clusters over here. AI clusters over there.”
Instead of “what would a bot write next,” it asks “who does this SOUND like?”
Two things you care about:
No single tell sinks you. Not the em dash. Not “delve.” It looks at hundreds of thousands of signals. So relax about your em dashes. (In theory.)
Length matters. More words = more accurate. That’s your 100-word floor.
Now… does it actually work?
The fine print nobody read
Okay here’s the part I’m actually fired up about.
July 20th , the day BEFORE the button launched , Substack quietly updated the Publisher Agreement. Former lawyer Noemi from ME TIME 🎈 brought this up today in our Club Chat.
The changelog literally says: refund tweak… and a “clarification regarding automated content analysis.”
You now grant Substack a license to run your content through automated tools, including AI detectors. So yeah, they can feed your writing to Pangram.
They say they can’t use your content to train third-party AI models, and Pangram can’t keep your content or use it for anything except the scan. So guardrails are baked in…
If a tool slaps a label on you that changes how your work shows up? They have to let you contest it. That’s the legal spine under the “report this scan” button.
Here’s the word I want you to circle: generative.
Their promise says they won’t use your stuff to train GENERATIVE AI.
But Pangram isn’t generative. It’s a classifier.
So the Club’s AI expert Patrick is asking:
“Does that promise cover training the detector (AI tool) itself?”
What do you say? Join our discussion. This is the forum.
The privacy change on Substack nobody emailed you about
Okay this one bugs me….Thanks to the Club’s legal advisor Noemi from ME TIME 🎈 who pointing this out! today in the Club CHat.
Adding Pangram = adding a new subprocessor. A new company that touches your data.
And Substack’s own agreement says they’ll notify creators about subprocessor changes and give you a window to object.
So how do you get notified?
You… email them. To ask. You have to opt in to being told….
Read that twice.
That’s… a great question.
Whatever you think about the scan button, a company sitting on everybody’s email telling you to email in to hear about a contract change? That’s a weird look. Substack should just answer this one.
I could go on for hours… but as always, wanted to wait some days before I’d open the forum to talk about Substack and AI.
Now it’s your turn:
What do you think about this new Pengram x Substack collab?
Have you tried it?
How are you faring with AI-written or AI-assisted posts?
♻️♻️Please share this so more join this forum to talk about Substack, it’s new AI tool and pdated Publisher Agreement♻️♻️ I’m writing this past midnight. I’m a part-time writer offering you waht I can for free.
A restack is sooo easy and it cost nothing. A comments let’s me keep going and if you become a paid sub… ohhh I do my happy dance and see you in our LIVES on Friday.
I'm incensed that they're being coy about the fact that Pangram gets paid per words scanned. That's the first reason everyone was livid that they were being forced to scan text to opt out.
I'm also angry about the fact that their scans are so faulty. Make grammatical mistakes or typos? 💯 human. Fix those and suddenly your text is 99% AI.
It's a false and misleading metric.
And if, as you pointed out at the start, it was already easy to detect that something was badly written AI...why do we need a machine to tell us something that we should be learning to figure out for ourselves?
Detecting human words from machine language is a skill we should be honing, not one we should be outsourcing to...machines.
Moreover... if the ideas are human inspired and the content interesting and relevant, who cares if an AI put the words together?
Ai isn't inherently evil. It's a tool. People who use it to spit out regurgitated content scraped from other sources or to make a quick buck off unsuspecting consumers are evil...Pangram isn't going to help us weed those out. It's just going to keep good writers from risking their reputations on a very fallible tool with shady business practices.