Substack's new AI scanner tells paying readers exactly how much of their newsletter a human actually wrote

Substack's AI Scanner Doesn't Police Writers — It Outsources the Verdict to Readers, and That's the Bigger Story

Substack switched on an AI-detection feature on July 21, 2026, built on third-party tool Pangram, that lets any subscriber privately scan a post, note, reply or comment for an estimated human-vs-AI-generated breakdown. Co-founder and CEO Chris Best introduced it in a post titled "Against Claudefishing," coining that term for newsletters that read as personal but were substantially AI-drafted. The tool is live on web and iOS, with Android still pending, works only on text published on or after July 21, 2026, and requires a minimum of 100 words to return a result. Scans stay private to the reader who requested them and don't cover video, audio, email, standalone Substack sites or custom domains. Substack takes 10% of paid subscription revenue on the platform, which is the commercial context for why authorship now gets checked.

What does the scanner actually show a reader?

A private percentage estimate of human-written versus AI-assisted text for a single post, note, reply or comment — nothing more, nothing public.

It's a per-reader lookup, not a badge, a public score, or a platform-wide ranking. A subscriber has to actively request a scan; nothing changes on the page for anyone who doesn't. That design choice matters for a niche audience used to detector output being misread as a public verdict — this one deliberately isn't.

How much should buyers trust Pangram's false-positive claim?

Treat "about one in 10,000" as a best case: Pangram's own blog says accuracy drops on short, formulaic or unusual writing.

That's the vendor stating the caveat itself, not a critic finding it. Outlines, lists, spreadsheets and instruction-manual-style prose are exactly the formats Pangram flags as weaker ground for its model — and exactly the formats a lot of paid newsletters use. Best has also been explicit that the tool can't distinguish AI-drafted text from a human who used AI as a source or heavily rewrote AI output afterward. Any reader treating a scan as proof of ghostwriting is over-reading a number the vendor itself qualifies.

What should a paid newsletter writer do differently now?

Assume any post can be scanned by a paying subscriber, and be ready to explain your process if the estimate looks high.

Pangram raised $9M in a round led by Menlo Ventures, taking its total funding to roughly $14M, alongside a new detection model release timed to Substack's rollout — a reminder that detection-as-a-service is now a funded category with its own incentive to show usage growing, not just accuracy. Writers who use AI for research, outlines or first drafts should keep that workflow documented, because "the scanner said X%" is now a conversation readers can start unprompted.

Will Beehiiv, Ghost and Kit have to match this?

Competitive pressure is real, but none has shipped a comparable reader-facing scanner yet, leaving Substack the only newsletter platform making this claim.

For buyers comparing platforms, the honest question isn't "who has AI detection" — it's whether a rival platform's silence on authorship checks is a considered choice or just a gap they haven't gotten to.

Frequently asked questions

Can writers opt out of being scanned?

No opt-out is described; any reader can request a scan on eligible posts, notes, replies or comments published after July 21, 2026.

Does this cover free posts, or only paid ones?

The source doesn't restrict the feature to paid content specifically — it applies to eligible text-based posts and notes generally, not just subscriber-only writing.

Is the scan result ever shown publicly?

No — results are private to the individual reader who ran the scan, not published to the post or visible to other subscribers.

Sources: Startup Fortune.