The platform built on human voices just admitted it can't tell them apart anymore.
The Summary
- Substack is rolling out an AI detection tool powered by a company called Pangram that scans posts, notes, replies, and comments for AI-generated text
- Readers can access the scanner from a three-dot menu on any content longer than 100 words to get an estimate of how much was written by AI versus humans
- The feature launches on web and iOS first, with Android "soon," signaling Substack sees enough AI slop in its ecosystem to build platform-level defenses
The Signal
Substack just built a triage tool for a problem it didn't create but can't ignore. The new AI scanner, powered by detection firm Pangram, gives readers a way to check if what they're reading came from a human or a language model. Click the three-dot menu, select "Scan for AI text," and get an estimate. It works on anything over 100 words: full posts, quick notes, comment threads.
The timing matters. Substack sells itself as the anti-algorithm, the place where real writers build direct relationships with real readers. But AI writing tools are everywhere now, and the platform's open model means anyone can spin up a newsletter, paste in Claude or ChatGPT output, and hit publish. Mashable notes the tool is explicitly designed to help readers figure out whether content was written by a human or generated by AI, framing it as a transparency play.
"Substack will now help users determine whether what they're reading may have been written by AI."
But here's the real tell: Substack isn't blocking AI content. It's labeling it. That's a choice. They could have banned AI-generated newsletters outright or required writers to disclose AI use upfront. Instead, they're putting the detection burden on readers and making it opt-in. You have to actively choose to scan. The default is trust.
That's not a technical limitation. It's a business decision. Substack takes a cut of paid subscriptions. If a newsletter grows an audience and converts readers to paid subscribers, the platform doesn't care if the prose came from a person or a prompt. What they do care about is keeping readers from feeling duped, because that's what drives churn.
Key dynamics at play:
- Detection is reactive, not preventive: the tool only works if readers suspect something and choose to scan
- Platform economics reward engagement and conversions, not authorship verification
- Substack is outsourcing the hard problem (determining AI use) to a third party, Pangram, rather than building in-house
The rollout starts on web and iOS, with Android coming later. That's standard product sequencing, but it also means a chunk of Substack's mobile readers won't have access to the tool at launch. During that window, AI-generated content flows unchecked to anyone reading on Android.
The Implication
If you're a Substack writer who uses AI to draft, edit, or speed up production, this won't stop you. But it will make some readers wonder. The smarter play: disclose your process upfront. Tell people if you're using AI for research, outlining, or drafts. The ones who care will appreciate the honesty. The ones who don't will keep reading.
For readers, the tool is a signal that the problem is real enough for Substack to acknowledge it. Start scanning the newsletters you pay for. If they're 80% AI and you didn't know, you're funding a content farm, not a writer.