Substack Adds AI Detection So Readers Can Check Who Actually Wrote a Post

You read a newsletter, something feels a little too smooth, and you can't quite put your finger on why. Substack is now handing readers a way to answer that nagging question instead of just sitting with it. The company has teamed up with AI-detection firm Pangram to roll out scanning tools that check posts, notes, and replies for signs of machine-written text.

CEO Chris Best laid out the thinking in a post he called "Against Claudefishing." The term is his own — it describes writing that leans on AI while quietly presenting itself as human work. That framing matters, because the target here isn't AI use in general. It's the pretending.

How Substack's AI Detection Feature Actually Works

The mechanics are refreshingly narrow, which is probably a good sign. Scans only run on text longer than a hundred words, and only on material published from July 21 onward. Shorter notes and older archives stay out of range.

Here's the part that changes the social dynamic: the result is private. Only the person who requested the scan sees it. There's no badge slapped onto a writer's post, no public scarlet letter, no comment section pile-on triggered by a score. You check, you get your answer, and that's the end of it.

Best isn't overselling the accuracy either. He's said plainly that Pangram gets things wrong sometimes, while pointing to independent evaluations that credit the tool with a high degree of accuracy. That's a more honest posture than most detection pitches manage.

The feature is live on web and iOS right now. Android users are waiting.

Where the Scan Works and Where It Doesn't

Why Substack Built This Now

The timing isn't accidental. Best points to a Pangram estimate suggesting that on some platforms, as much as forty percent of posts are now fully AI-generated. Forty percent. That's not a fringe problem creeping in at the edges — that's a substantial chunk of what people scroll past every day.

For a platform built on the premise that you're paying to hear from a specific person with a specific mind, that number is closer to an existential threat than a content-moderation headache. Subscriptions are a bet on a writer. If the writer turns out to be a prompt, the bet was never real.

What Writers Get: Transparency Instead of Suspicion

The genuinely interesting move is that Substack didn't stop at giving readers a detector. Writers get tools too, and they change the tone of the whole thing.

There's a new "How I make this" statement, where writers describe their own process and set expectations before anyone starts guessing. Some people outline with AI and write by hand. Some draft clean and use a model to catch typos. Some don't touch it at all. Saying so upfront is a lot more useful than leaving readers to squint at sentence rhythm.

Writers can also run Pangram on their own drafts before hitting publish. That's a quiet acknowledgment of something real: a false positive can hit an honest writer just as hard as a true positive hits a dishonest one. Being able to check your own work first turns a gotcha tool into something closer to a spell-checker.

And if a scan of published work looks wrong, writers can report it and have it removed. The appeal path exists from day one, not as a patch after the first controversy.

What Might Come Next

Best has floated a few directions the feature could grow. AI preferences could land inside Reply Rules, letting writers set terms for what shows up in their comments. Readers might get controls over what gets recommended to them. Neither is shipped — both suggest Substack sees detection as one piece of a larger system rather than a finished product.

Substack Isn't Alone in Chasing AI Slop

Other platforms are circling the same problem with different tactics.

• LinkedIn has begun cutting the reach of AI-generated posts and comments, treating slop as a distribution problem rather than a labeling one.
• Meta is quietly working on its own detection tool, though nothing has gone live yet.

Substack's version stands out for one reason: it gives the reader the button. LinkedIn's approach adjusts what you see without telling you why. Meta's isn't in anyone's hands. This one puts the decision with the person actually reading.

The Question Nobody Can Answer Yet

Detection tools and generation models are locked in the same race, and the models are the ones getting a fresh release every few months. Whether any of these systems can keep up with what they're built to catch is genuinely unclear.

But there's a case that keeping perfect pace isn't the point. Making deception slightly inconvenient, slightly riskier, slightly more likely to get noticed — that shifts incentives even when the tool misses. A writer who knows readers can check is a writer thinking twice.

For readers, this is a rare thing: a transparency feature that hands you an actual control instead of a promise. Use it when something feels off. Ignore it when it doesn't. Either way, the choice finally sits with you.