Substack has added an on-demand scanner that estimates whether artificial intelligence contributed to text published on its platform. The Pangram-powered feature covers newsletters, Notes, replies and comments, giving readers a direct check instead of placing an automatic warning on every piece of suspected writing. That choice fits a service built around named authors and subscriptions, where confidence in who produced a post can influence whether a reader pays or returns.
The rollout began on July 21, 2026, across Substack's website and iOS app; Android support is due later. A reader opens the three-dot menu on eligible content and selects the scan option. The text must exceed 100 words, and the feature applies to material published from the launch date rather than serving as a universal audit of the archive. The length floor excludes short reactions, captions and one-line replies that would give the classifier too little material.
Pangram returns an assessment of how much writing may have been generated or assisted by AI. That wording is important because the software is classifying patterns, not retrieving a reliable record of which tools an author used. Substack is presenting the output as information for the reader, leaving the decision to trust or ignore it with the person viewing the post. It is neither an authorship certificate nor a record drawn from the writer's account activity. A precise-looking estimate can still convey more certainty than the underlying inference deserves.
Writers Get a Disclosure Box and a Way to Appeal
Authors receive tools on the other side of the same system. They can scan a draft before publication, report a result they believe is inaccurate and add a statement describing their production process. The disclosure space allows a writer to distinguish research, editing, transcription or image work from generated prose without forcing every use of software into one label. It also gives readers an explanation tied to the author rather than expecting a detection percentage to carry the entire judgment.
Substack CEO Chris Best framed the policy around a mismatch between what readers expect and what an author actually delivers. He calls undisclosed synthetic attempts at human connection “Claudefishing,” but does not treat all AI use as misconduct. Under that approach, a carefully edited piece using a model for limited assistance can be disclosed differently from an automated post presented as a personal account. The policy judges concealment and reader expectation separately from the literary quality of the output.
The detector cannot establish how much human judgment went into a finished article, and it cannot tell whether an AI tool merely supplied background information. False positives can place an honest author under suspicion; false negatives can give mass-produced text a clean result. Appeals are therefore essential, but Substack has not described a formal adjudication process or promised that a disputed score will be removed within a set period. An author's process statement may also contradict the scanner, so the interface must distinguish a personal disclosure from a classifier's claim. Public confidence will depend on whether challenged results lead to visible corrections rather than disappearing into a generic report queue.
A Reader Tool Shifts the Trust Dispute Into Every Post
Putting the scan command in the reader menu changes the social dynamic around detection. Authors no longer control when their work is tested, while commenters and newsletter writers can face public judgments based on a probabilistic output. The feature may deter undisclosed automation, yet it can also encourage readers to treat a software estimate as proof when vocabulary, translation or heavy editing produces an unusual pattern. Because the scan is voluntary, two readers may leave the same post with different levels of suspicion and different information.
Substack is betting that visible uncertainty is preferable to silence about authorship. That bet succeeds only if the company publishes error information, handles challenges quickly and prevents detection scores from becoming a shortcut for harassment. Writers also need to know whether a reported mistake affects distribution, subscriptions or moderation before they can judge the risk of opting into the system. Without those safeguards, the scanner will identify some hidden automation while manufacturing a second problem: writers forced to defend authentic work against a machine's unverified suspicion.