Substack has integrated Pangram, an AI writing detection engine, into its app, giving users the ability to scan posts, comments, notes, and replies for AI-generated content. The feature went live in July 2026 and is available for any post, note, reply, or comment exceeding 100 characters. (TechCrunch)

The detection tool provides an estimated breakdown of how much of a given piece was written by a human versus generated by AI — think of it as a percentage split rather than a binary verdict. It is not designed to block or penalize AI-assisted writing. Instead, Substack is positioning it as a transparency mechanism: writers can add an optional "AI Author's note" disclosing their use of AI, and the platform is encouraging creators to explain their process through a "how I make this" statement. (TechCrunch)

Publishers retain control over how the tool interacts with their own work. Writers can run Pangram on drafts before publication to self-check, and they can report and request removal of scans on their own content if they believe the detection result is incorrect. Substack CEO Chris Best described the feature as "good use of AI." (TechCrunch; Substack)

The design choices here matter. Substack has built a platform whose value proposition to writers and readers alike is direct, unmediated relationships. Bringing in AI detection at the reader-facing layer is a different posture from, say, an academic integrity tool that lives inside a submission pipeline. Readers can run a scan on someone else's newsletter post in real time, inside the app, and see a human-to-AI ratio estimate. That is a structural choice with implications for trust between writers and their audiences.

The decision not to penalize AI-assisted writing fits with how Substack has handled content governance broadly. The platform has generally favored disclosure and reader choice over hard enforcement. The Pangram integration extends that philosophy: the tool surfaces information and lets readers decide what to do with it, rather than imposing a prohibition. The optional "AI Author's note" and the "how I make this" framing put the onus on writers to be transparent, but do not require it.

Whether Pangram's detection accuracy holds up under scrutiny is an open question. AI detection tools across the industry have a well-documented track record of false positives, particularly with non-native English writing and certain stylistic patterns. Substack mitigates this by giving publishers the ability to contest scans on their own work, but the reader-facing nature of the feature means that an inaccurate detection result could influence reader perception before a writer even knows a scan has been run. In my view, that asymmetry — readers can scan a writer's work without notification, but writers must actively discover and contest false results — is the most consequential design tension in the integration, and it will be worth watching how it plays out in practice.

The 100-character threshold is a practical floor. Below that, detection signals are too noisy and the analytical value is marginal. Above it, Pangram has enough text to produce a meaningful estimate. Notes and replies are included alongside full posts, which means the feature covers Substack's shorter-form social layer, not just long-form newsletters.

The broader context is that AI-generated content has become pervasive across publishing platforms over the past two years, and platforms are now working through how to give audiences signal about what they are reading. Substack's approach — detection plus optional disclosure, no prohibition — is one model. It treats transparency as a reader-side tool and writer-side responsibility, rather than a platform-enforced rule. Whether that model proves durable depends on whether detection accuracy keeps pace with generative model quality, and whether readers actually use the information to make different choices about what they subscribe to and support.

For writers who use AI tools in their workflow, whether for drafting, editing, research synthesis, or translation, the integration introduces a new variable. A reader running a Pangram scan on a heavily AI-assisted post that carries no disclosure could create friction that the writer did not anticipate. Conversely, writers who proactively disclose their process may find that transparency itself becomes a trust signal in an environment where AI provenance is increasingly scrutinized.

Substack's bet, implicitly, is that readers want this information and that writers will adapt to its availability. The platform is not taking a side on whether AI-assisted writing is good or bad. It is providing a tool and letting the ecosystem negotiate around it.