Substack brings Pangram AI scans to posts, Notes and comments

Why it matters

Substack earns money when readers trust individual writers enough to pay. AI provenance is becoming part of that transaction, and Substack is testing disclosure before enforcement.

Substack CEO Chris Best launched a reader-triggered AI detector on July 21st, giving users a way to estimate how much of a post, note, comment, or reply was written by a person or with artificial intelligence.

The feature, called "Scan for AI text," integrates technology from Pangram, a Brooklyn AI-detection startup. TechCrunch reported the launch on July 22nd.

For Best, the product turns one of Substack's founding promises into a platform policy. He co-founded Substack in 2017 with journalist Hamish McKenzie and engineer Jairaj Sethi after serving as co-founder and CTO of messaging app Kik. Substack entered Y Combinator's Winter 2018 batch with a model built around readers paying writers directly.

Best's pitch was that software should handle the mechanics of publishing while writers supplied the thought readers valued. In a July 21st conversation with Pangram co-founder Max Spero, Best returned to that division of labor: "software should do everything else, but I think you do want the person to do the hard part."

The detector is Best's attempt to defend that hard part as AI makes authorship harder to establish.

A transparency tool with an escape hatch

According to Substack's announcement, readers can request a Pangram analysis instead of seeing AI labels placed automatically on every piece of content. The result estimates how much text was human-written or AI-assisted, and it is shown only to the person who initiated the scan.

The feature applies to eligible content published on or after July 21st. It is available on the web and iOS, with Android support listed as coming later. Substack's help documentation says scans work in the Substack Reader, the iOS app, Notes, comments, and replies.

Substack's help materials say the tool works on text above 100 characters, and short Notes may produce a "not enough text" response without specifying more detail.

Creators can scan drafts before publishing, attach a "How I make this" statement explaining their process, and report results they believe are wrong. They can also disable detection on their own posts and Notes. Readers then see that AI detection is unavailable rather than receiving a score.

That opt-out keeps Substack from turning a probabilistic classifier into an enforcement system. It also limits what readers can learn from the feature. A writer who objects to a result can remove access to it, leaving the audience to interpret the absence of a score.

The detector has another material boundary: it does not work in emails, audio or video posts, standalone Substack sites, or publications viewed through custom domains. Many subscribers still encounter newsletters in their inboxes, where Substack cannot place the same interactive controls around the text.

The rollout therefore has its greatest force inside the Substack app and Reader, where Substack owns the interface. That fits the direction Best and his co-founders laid out after Substack's $100 million Series C in July 2025, led by BOND and The Chernin Group. Substack said then that it would invest further in its app and network, expanding beyond the original email newsletter product.

Paid writing depends on provenance

Substack's economic exposure explains the timing. Publishing remains free, but Substack takes 10% of paid subscription transactions, before Stripe's processing charges. Substack benefits when readers believe a publication contains work worth paying for month after month.

Cheap, undisclosed AI production weakens that proposition. A reader paying for a writer's analysis, reporting, or personal voice may value the work differently after learning that a model generated much of it. Best calls that mismatch "Claudefishing," a term for presenting machine-produced work under an implied promise of human authorship.

Best is careful to leave room for authors who use AI as an editing, research, or production tool. Substack itself uses AI for software development, research, clipping, and translation. In the launch post, Best wrote that "people should know what they're getting."

That distinction gives creators latitude over their workflow while asking them to set accurate expectations. It also avoids treating every detected edit as evidence that a publication lacks human judgment. Pangram can estimate whether AI touched the prose. It cannot establish who developed the idea, conducted the reporting, rejected a model's mistakes, or accepted responsibility for the final argument.

Pangram gets a distribution test

The integration gives Pangram a prominent consumer deployment outside its early foothold in education and enterprise review systems.

Pangram co-founders Max Spero and Bradley Emi met as Stanford undergraduates before working on applied machine learning. Spero worked at Nuro, Google, Two Sigma, and Yelp. Emi worked on Tesla Autopilot and later led deep-learning research at drug-discovery startup Absci.

Pangram announced a $4 million seed round in June 2025.

Substack gives Spero and Emi a prominent test across professional journalism, hobbyist posts, academic writing, comments, and multilingual writing. Creator error reports create a feedback channel for the product, while every mistaken classification risks damaging trust in both products.

Substack has not published performance data for this particular integration. The help materials do not provide false-positive rates, language-by-language results, or details about how reported errors are reviewed.

The concern extends beyond Pangram. A 2023 Stanford study found that several GPT detectors frequently misclassified writing by non-native English speakers. That research did not evaluate the current Pangram integration, but it established the cost of presenting detection scores without sufficient context.

Best has chosen a restrained first deployment: users must ask for a scan, authors can explain their methods, and no automatic penalty follows a result. The design treats detection as evidence for a reader's judgment rather than a verdict on a writer.

Substack's business was built on the idea that a direct financial relationship can align writers with readers better than advertising-driven distribution. AI adds a new condition to that relationship: readers need some basis for knowing whose judgment they are purchasing. Best is betting that disclosure and optional detection can supply it without dictating how writers work.