Substack and Pangram team up to fight "Claudefishing"

You finish reading a moving essay, connect deeply with what felt like raw human emotion, and then realize it was prompted in thirty seconds. Substack now has a name for that betrayal: Claudefishing.

The publishing platform launched a partnership with AI-detection company Pangram, giving readers and creators built-in tools to spot automated writing across the network.

Scanning for ghostwriters

Pangram estimates that AI-generated text already makes up as much as 40% of posts on some social platforms. Substack CEO Chris Best made it clear the company doesn't want to wait until its feed turns into LinkedIn before pushing back.

Starting today on web and iOS, users can run an on-demand scan on any post, Note, reply, or comment longer than 100 words published from today onward. The scan estimates how much of the piece was written by hand versus produced with AI assistance.

Crucially, scans don't show up automatically. Analyses only appear when an individual reader requests one. Substack also noted that Pangram isn't flawless: it detects whether language was generated by AI, not how much thought went into prompting or editing, nor whether AI was merely used as a research source.

A nutrition label for creators

To protect writers from false accusations, Substack is giving creators some defensive gear as well.

Writers can run Pangram scans on drafts before hitting publish, and they can report or remove scan results on their own work if they believe the tool got it wrong. Substack is also rolling out a "How I make this" statement, letting authors explicitly declare their process and set clear expectations with subscribers.

An Android release is coming soon. The platform is also considering future options like letting writers ban AI replies in their community rules and giving readers preferences over AI-assisted content in recommendations.

Why it matters

Substack runs on reader trust. People pay subscriptions because they care about a specific person's perspective and craft, not because they want machine-generated filler. When readers feel tricked by a synthetic voice masquerading as a human, the economic engine collapses.

We think Substack handled the rollout smartly by keeping scans opt-in rather than slapping automated warning badges across every article. Letting creators publish their own workflow disclosures treats readers like adults who can weigh assistance against authenticity.

The obvious trap is that AI detectors are notoriously imperfect. Even top-tier detection engines struggle with false positives on dry, formulaic human writing, while savvy prompters can easily bypass them. If readers start weaponizing flawed scan scores in comment sections, Substack might end up manufacturing the very paranoia it set out to cure.