Substack launches Pangram AI detector to label AI-generated content

Substack has integrated the Pangram AI detector into its platform, giving readers a way to scan posts, notes, replies, and comments for AI-generated content. The tool, available on web and iOS with Android coming later, labels text as AI-generated, AI-assisted, or human. For AI builders and content creators, this signals a growing expectation for transparency around AI-assisted writing and could influence how platforms handle content provenance.

What happened

Substack partnered with Pangram to deploy a classifier neural network trained on pre-2021 human text to distinguish human from AI writing. Readers access the tool via the "Scan for AI text" option in the three-dot menu. It only works on passages over 100 words published after 8:30 a.m. PT on July 21. Authors can add a statement explaining their writing process, run the detector on drafts, submit correction reports if misclassified, or disable AI detection entirely. Substack CEO Chris Best noted the tool cannot detect AI used for research before human writing. Substack may add more AI features based on feedback, including content-filter preferences.

Why AI builders should care

This feature represents a platform-level signal for content provenance. For builders of AI writing tools or platforms that host user-generated content, similar transparency requirements may become standard. The ability for authors to disable detection also raises questions about trust signals and how platforms balance transparency with author control. The Pangram detector's reliance on pre-2021 training data means it may miss newer AI writing patterns, which is a caveat for anyone relying on such tools.

Practical implications

Authors can now proactively disclose their use of AI by adding a statement visible when readers scan. They can also pre-scan drafts and contest false positives. The option to disable detection entirely means some content will remain unlabeled, shifting the burden of provenance verification to readers. For builders integrating AI detection into their own products, Pangram's approach of using a classifier rather than generative AI signals a design choice that avoids some contamination issues but may have blind spots.

Caveats

The detector is not universal. It only applies to content longer than 100 words published after the cutoff date. It cannot detect AI used for research or editing. Misclassifications are possible, and authors can disable detection, which undermines the tool's reliability as a definitive signal. Pangram's training data stops at 2021, so newer AI writing styles may not be recognized. Readers must verify provenance themselves; the tool is a signal, not a verdict.