"Claudefishing": Substack Gives Readers a Button to Uncover AI-Written Posts

Substack has coined a term for the moment you discover that a post admired for its voice was actually produced by a bot: Claudefishing. Not only the name, but also a button to report it: the newsletter platform has launched an AI detection feature developed with Pangram, introduced on Tuesday in a post explicitly titled "Against Claudefishing."

From the voice "Scan for AI text" in the three-dots menu of a piece of content, the reader analyzes posts, notes, replies, and comments to get an estimate of how much of the text was written by hand and how much with AI assistance. The scan works on texts longer than 100 words published from Tuesday onward, and the result is visible only to those who request it. The feature is already active on web and iOS, with Android coming soon.

Against Claudefishing by Substack

Building trust in the AI age Read on Substack

Substack clarifies that the function does not aim to prohibit or penalize assisted writing. Authors can run the detector on their drafts before publishing, add a "How I made this" statement to explain their process, and report or remove analyses of their work that they consider incorrect. The company itself uses AI first-hand to write software, do research, and build product features like translations.

The Target Is the Gap Between Expectations and Reality

Co-founder and CEO Chris Best is clear on the point: "The underlying problem is not that people use AI, nor the quality of what it produces. It arises when there is a gap between the reader’s expectations and reality, especially when someone invests their attention, unknowingly, in something that has no human thought behind it. This is Claudefishing." The closing remark is blunt: "Claude has much to offer. But when I want Claude's opinion, I ask him directly."

The post does not hold back from a jab at LinkedIn, indicated by many as the epicenter of machine-generated prose. The stakes are high for Best: when readers must wonder if what they are reading is real, "trust in authorship erodes and jeopardizes the livelihood of those who write." Pangram estimates that up to 40% of text on some social platforms is now generated by AI.

Detection Is Not Established Science

AI detection remains uncertain territory, and it is here that Substack acknowledges its own limitations: Best himself admits that Pangram "can only detect if AI was used to produce the text, not if there was human care behind it." The Atlantic has measured how accurate these tools are, including Pangram, with a warning: trusting too much risks turning accusations into a witch hunt.

Any oracle for "written by an AI" can be used to fine-tune an AI to evade the detector. There is no point to such detectors; they cannot work for long, certainly not as commercial products. https://t.co/VzhHwTJI1K — Perry E. Metzger (@perrymetzger) July 21, 2026

Developer Perry Metzger has noted that any detector allows tuning an AI capable of evading it, making such tools short-lived. Researcher Mor Naaman has referred to a lost battle: if readers scan, writers will do so too, bypassing the control. As one commentator succinctly put it under Substack's post, the presence of AI does not prove the absence of a human.