Substack has launched an AI-generated text detection tool powered by Pangram, giving readers and writers a way to scan posts, notes, replies, and comments for AI involvement. The feature was announced in a Substack blog post on July 21, 2026 (The Verge).

The tool is rolling out on web and iOS, with Android support stated as coming "soon." Readers can analyze any Substack content longer than 100 words by selecting "Scan for AI text" from the three-dot menu in the top-right corner of a post. Writers can also scan their own drafts and have an option to report inaccurate results.

Substack co-founder and CEO Chris Best said the Pangram integration is meant to increase transparency as AI becomes more prevalent on social media. Best coined the term "Claudefishing" to describe the mismatch between a reader's expectation of human authorship and the AI-generated reality of what they are reading. The concept is outlined in Substack's own post on the subject (Substack).

Alongside the detection tool, Substack is launching a "How I make this" statement feature that allows creators to explain their writing process to readers. The two features together form a transparency package: one lets readers probe text for machine involvement, the other lets writers proactively disclose their methods.

Best was candid about the limitations. He stated that Pangram can only detect whether AI was used to make the text, not whether great human care went into creating it or whether AI tools were used as a source. That is a meaningful distinction. A writer who drafts in their own words but uses an LLM for research, fact-checking, or structural suggestions may still trigger a positive detection, while a heavily AI-assisted piece that has been paraphrased or rewritten by a human may not. The tool addresses provenance, not effort or craft.

Pangram, which operates at pangram.com, brings genuine technical credentials to the partnership. The company's latest detection model, version 3.3, was detailed in a July 9, 2026 blog post (Pangram). Pangram claims version 3.3 achieves a 0.01% false positive rate in AI detection. If that figure holds in production conditions across Substack's heterogeneous content base, it would place Pangram among the more conservative detectors on the market; false positives have been the central failure mode of AI detection tools since the category emerged, and a low FP rate is the metric that matters most for a platform deploying detection at reader-facing scale.

Pangram's detection technology has broader reach. The company offers a Chrome extension, marketed under the name "Feed Scanner," that can detect AI-generated content on X, LinkedIn, Substack, Medium, Reddit, and Google Docs (Pangram). Separately, Pangram's technical manuscript "EditLens" was accepted for publication at ICLR 2026 (Pangram), giving the company a peer-reviewed foothold in the machine learning research community.

The broader context here is the tension between detection accuracy and platform dynamics. Substack's model is built on direct subscriptions and reader trust in individual writers. AI-generated content flooding a subscription feed erodes that trust at the foundation, which is presumably why Best has framed the problem around the reader-writer relationship rather than around content moderation in the traditional sense. The platform is not removing or downranking flagged content; it is giving readers the means to make their own judgments.

Worth flagging is the gap between a 0.01% false positive rate in a benchmark and the realities of deployment across a platform where writing styles range from casual newsletter posts to long-form investigative journalism. Pangram's claimed figure is a laboratory result; how it performs when applied to Substack's full content corpus, including non-native English speakers and stylistically unconventional writers, will determine whether the feature earns trust or generates friction. The in-product "report inaccurate results" option is an implicit acknowledgment that the system will produce errors in the wild.

The "How I make this" feature may ultimately carry more weight than the scanner itself. Voluntary disclosure shifts the transparency question from a technical arms race between generators and detectors to a social norm around expectations. If enough writers adopt it, readers calibrate their own trust rather than relying on a probability score. If adoption is low, the scanner becomes the primary signal, and the burden falls entirely on Pangram's model.

For now, Substack has placed a detection tool and a disclosure mechanism side by side, betting that the combination gives readers enough signal to navigate an increasingly synthetic content landscape.