Substack ships AI detection to smoke out 'Claudefishing' newsletters

Substack rolled out a new Scan for AI text feature on July 21, 2026, powered by the third-party detection company Pangram. According to Substack's official support documentation, the tool works on posts and notes published on or after that date, and it assigns a percentage estimate of how much content was written by a human versus AI-assisted. The scan is available on the Substack Reader web app and the iOS app, with Android support confirmed as coming soon. Readers can also run the scan on comments, provided the text exceeds 100 words.

The feature sits alongside a new disclosure space where creators can explicitly state if and how they used AI in their work. Substack's blog post, titled Against Claudefishing, frames the tool as a transparency measure rather than a purity test. CEO Chris Best qualified that not all AI writing is slop, but the real problem emerges when authors misrepresent machine-generated prose as their own labor. The term Claudefishing, a coinage targeting Anthropic's Claude model specifically, captures the practice of passing off AI drafts as human effort to build subscriber trust on false pretenses.

Response to the announcement was immediate and polarized. Theaugmentededucator published notes on bypassing Pangram detection within hours, treating the tool as a challenge rather than a deterrent. Ruben.substack went further, calling AI detection bots a scam outright and arguing that human readers already possess better intuition for spotting synthetic prose, pointing to telltale vocabulary like delve, tapestry, and realm, plus the overuse of negation framing patterns. Giuliacassara.substack offered a qualitative guide for detecting AI-assisted articles without relying on automated tools, describing the experience of encountering a Substack post that felt like every paragraph was a nuanced repetition of the main argument, overexplaining a concept that could have been just 50 words.

This skepticism mirrors longstanding doubts about AI detection accuracy. Sf.gazetteer noted Pangram is imperfect but industry-leading, a diplomatic way of acknowledging the fundamental difficulty of reliably distinguishing human and machine text. The tool's deployment on a platform built around paid subscriptions raises the stakes: a false positive could damage a writer's reputation and income, while false negatives let Claudefishing continue under the radar.

TechCrunch reported that the move might be bad for Substack's business in the short term, since it could expose many newsletters that rely on AI assistance. A Wired analysis from November 2024 estimated that roughly 10 percent of Substack's top newsletters already publish AI-generated or AI-assisted content. Those top writers earn up to seven figures annually through paid subscriptions, creating a business model with different incentives than ad-driven platforms like Facebook or YouTube where engagement metrics reign supreme.

Substack is effectively betting that reader trust matters more than protecting the revenue streams of AI-reliant creators. If the detection tool surfaces widespread AI use among popular paid newsletters, subscribers may reconsider where they spend their money. The platform's subscription-driven model means transparency cuts both ways: it can strengthen loyalty when human authorship is confirmed, but it can also trigger cancellations when readers feel they've been paying for machine output marketed as personal correspondence.

Substack's integration is not an isolated experiment. It reflects a growing industry push for disclosure around AI-assisted content across platforms where trust is the core product. The new creator statement field lets writers voluntarily declare their AI usage, creating a norm where silence on the topic itself becomes a signal. TechCrunch characterized the launch as signaling a broader shift toward transparency around AI-assisted content, and the timing aligns with mounting reader fatigue over undifferentiated AI prose flooding social platforms.

Roberthiett.substack framed the moment as a call to work with AI responsibly, acknowledging that AI-generated content is everywhere, from blog posts to social media updates and emails, and that sometimes it's painfully obvious while other times it's nearly indistinguishable from human writing. Substack's dual approach, combining automated scanning with voluntary disclosure, attempts to address both ends of that spectrum: the blatant cases Pangram can flag, and the subtler uses that only the author can honestly contextualize.

AI detection remains a technically unsolved problem, and Substack's implementation inherits all of its limitations. Pangram provides percentage estimates, not definitive verdicts, and every source covering the launch acknowledged the imperfection. Ruben.substack argued that if humans can intuitively detect AI text through stylistic tells, a bot should be able to do it too, but the reality is more complex: detection tools produce probabilistic guesses that can be gamed through minor edits, prompt engineering, or simply writing in a voice that doesn't trigger the classifier.

Theaugmentededucator's immediate publication of bypass techniques underscores how quickly detection arms races escalate. For Substack readers, the practical implication is that the scan button offers a directional signal, not a courtroom-grade determination. A high AI percentage should prompt closer reading rather than an instant unsubscribe, and a low percentage doesn't guarantee human authorship any more than a polygraph guarantees truthfulness.

Substack confirmed Android support is in development, extending the detection tool's reach beyond web and iOS. The feature's real impact will emerge over months as scan data accumulates and patterns become visible. Writers who depend on AI drafting face a choice: disclose their workflow proactively, modify their process to evade detection, or risk reader backlash when scans reveal high AI percentages on paid content. Readers gain a new lens for evaluating where to direct their subscription dollars, but they also inherit the responsibility of interpreting imperfect probability scores without overreacting.

The creator disclosure field may prove more durable than the automated scanner. Voluntary transparency norms, once established, are harder to game than detection algorithms because they shift the social cost of deception onto the deceiver. If Substack's top earners embrace disclosure, it could set an expectation that trickles down through the platform's entire writer ecosystem, making Claudefishing socially risky even when technically undetectable.