Substack just switched on a feature that scans any post, note, reply, or comment on its platform and estimates how much of it was written by AI. The tool comes from a new partnership with Pangram, an AI detection startup, and it works on anything over 100 characters. If you write or read newsletters, this changes what “trust the byline” actually means.
What Substack Actually Turned On
A few days ago, Substack announced an integration with Pangram, an AI writing detection company, that lets anyone in the Substack app scan a piece of writing and get back an estimate of how much of it was human and how much was AI. It works on posts, notes, replies, and comments, as long as the text runs longer than 100 characters. That is a low bar. Most comments clear it without trying.
Writers also get an optional AI author’s note, a spot to disclose how they used AI in putting a piece together. Substack has been clear that the feature is not a ban or a penalty system. Publishers can run Pangram on their own drafts before hitting publish, and if a scan flags something they believe is wrong, they can report it and have it removed from their own work.
Substack CEO Chris Best framed the pitch to writers in a conversation with Pangram co-founder Max Spero: the platform handles the software, but the writer still has to bring something worth reading. Best called it “good use of AI,” which is a fairly low-key way to describe a feature that could expose a decent chunk of the platform’s content.
Why Now, And Why This Isn’t Happening In Isolation
Substack didn’t invent this idea. It’s catching up to a pattern that’s been building across nearly every content platform for the past two years, as AI-generated material got harder to spot with the naked eye and platforms started feeling pressure to help readers tell the difference.
| Platform | What it does | Consequence for AI content |
|---|---|---|
| TikTok | Auto-labels AI-generated video and photo content | Label added, content stays up |
| YouTube | Automatically labels AI-made videos | Label added, content stays up |
| Spotify | Labels AI vocals and flags spam patterns | Label added, repeat spam can be removed |
| Tidal | Detects AI-generated music | Monetization cut off |
| Substack | Scans posts, notes, replies, and comments for AI authorship | Estimate shown, no penalty, optional disclosure |
Notice where Substack lands on that list. It’s the only one of the five that explicitly says the goal isn’t to punish AI use, just to make it visible. Whether that holds up once readers start actually seeing the scores is a different question.
How To Actually Use The Scanner
If you use Substack, either as a reader or a writer, here’s what to do with this feature right now.
As a reader
- Open any post, note, reply, or comment in the Substack app that’s longer than 100 characters.
- Look for the scan option tied to that piece of content. It returns a human versus AI estimate rather than a flat yes or no.
- Treat the score as a signal, not a verdict. Detection tools are probabilistic. A high AI estimate on a short comment carries a lot less weight than the same score on a 2,000-word essay.
- Check for an AI author’s note on the writer’s profile or post before assuming anything. Some writers will disclose their process upfront, which makes the scan mostly redundant for their content.
As a writer
- Run your own draft through Pangram before you publish it, especially if you used AI for research, outlining, or a first pass. Knowing your own score before a reader finds it first puts you ahead of the conversation.
- Add an AI author’s note if your process involves AI in any real way. A short, honest line, something like “drafted with AI assistance, edited and fact-checked by me”, costs you nothing and heads off the awkward conversation later.
- Report a scan you believe is wrong. Substack allows publishers to flag and remove scans on their own work if they think the tool got it wrong. Detection models misfire on formulaic writing, non-native English phrasing, and heavily edited text, so don’t assume a flag is automatically correct.
- Rewrite the parts that read like a template, not to game the detector, but because “written to sound human” and “written to avoid detection” tend to be the same skill anyway. Clean, specific, opinionated writing scores lower on AI detectors almost as a side effect.
What This Means If You Run A Newsletter
The short-term risk for Substack is obvious. A lot of newsletters on the platform lean on AI more than their bylines suggest, and a visible scanner is going to surface some of that whether writers want it surfaced or not. That’s an uncomfortable moment for a platform whose entire brand rests on “independent writers, real voices.”
The longer game is different. If readers start trusting that a Substack byline means something, actual human effort behind the words, that trust becomes a competitive advantage over platforms where nobody’s even asking the question. Substack is betting that transparency now beats a slow erosion of trust later, once readers figure out on their own how much AI-slop has been quietly filling their inboxes.
If you write a paid or free newsletter anywhere, not just on Substack, this is worth getting ahead of. Readers are getting more literate about spotting AI patterns, tools like Pangram are getting cheaper and more accessible, and disclosure is becoming the norm rather than the exception. Deciding your own AI disclosure policy now, before a reader or a platform forces the question, is the cheaper version of this problem.
The Limitations Worth Knowing
AI detection is not a lie detector. Pangram, founded in 2023 by Stanford classmates Max Spero and Bradley Emi, has built its reputation on being more accurate than older tools like GPT-2 Output Detector or Turnitin’s AI checker, but “more accurate” is not the same as “always right.”
A few things to keep in mind before treating any score as gospel:
- Detection models can misread heavily formulaic writing as AI-generated, even when a human wrote every word.
- Very short text, comments especially, gives the model less to work with, so confidence drops even though a number still shows up.
- Mixed authorship, a human draft cleaned up by AI, or an AI draft heavily rewritten by a human, is genuinely hard to score cleanly. Pangram’s own team has talked about running a first pass to find rough boundaries and a second, finer pass for exactly this kind of blended text.
- None of this stops someone from disputing a fair flag just as easily as an unfair one. A dispute button doesn’t automatically mean the original score was wrong.
Bottom Line
Substack just made AI use in newsletters visible instead of invisible. That’s a bigger deal than the feature itself sounds, because visibility is what every other content platform has been inching toward for two years, one label at a time. Whether it helps Substack’s brand or exposes a problem it would rather have left alone probably depends on how many of its writers were quietly leaning on AI more than their readers assumed. Either way, disclosure is no longer optional cover. It’s becoming the baseline.
If your inbox is full of newsletters you’ve never actually questioned, this is a good week to start asking who’s really writing them.
Sources
- TechCrunch (July 22, 2026), “Substack’s new tool tells you who’s been writing their newsletters with AI.” Link
- Substack, official announcement on X. Link
- Pangram Labs, official site. Link
- Pangram Labs, “About Us.” Link
- Chris Best, online chat with Max Spero via Substack Notes. Link
- TechCrunch (May 27, 2026), “YouTube will now automatically label AI videos.” Link
- TechCrunch (May 9, 2024), “TikTok will automatically label AI-generated content created on other platforms.” Link
- TechCrunch (September 25, 2025), “Spotify updates AI policy to label tracks, cut down on spam.” Link
- TechCrunch (June 29, 2026), “Tidal cracks down on AI music by cutting off monetization.” Link
Just Being Resourceful! 
Leave a comment