Substack is adding an AI-text detector to its publishing platform, giving readers a built-in way to check whether posts and community writing may have been generated or assisted by artificial intelligence. The company described the feature in a Tuesday blog post by co-founder and CEO Chris Best.
The detector comes from Pangram, an AI detection company. According to Substack, it will work across posts, notes, replies and comments. Readers can use it on text longer than 100 words by opening the three-dot menu at the upper right of a post and selecting “Scan for AI text.”
Substack said the feature is rolling out on the web and in its iOS app. The company said Android support is coming “soon,” without giving a date.
The result is framed as an estimate, which is the sane way to talk about AI detection. These tools do not read a writer’s mind, inspect a draft history or prove intent. Substack says Pangram’s system can identify whether AI was used in making the text. Best also wrote that the tool does not determine whether a person put substantial care into the work, and it does not say whether AI tools were used only as a research source.
That caveat matters because Substack is not just adding a badge machine for dunking on bad prose. Best cast the move as a transparency feature aimed at a specific mismatch between what readers think they are getting and what they are actually reading. In his blog post, he used the term “Claudefishing” for cases where readers spend attention on writing that appears to have human thought behind it but may not.
Best wrote that the issue, in Substack’s view, is not the use of AI by itself or even whether its output is good. The problem is disclosure: readers may expect a person’s judgment, reporting or argument, while the text in front of them may have been produced with little human involvement.
Substack is also giving writers a way to describe their own process. The company said creators will be able to add a “How I make this” statement for readers. Writers can scan drafts with Pangram before publishing, and Substack said they will be able to report results they believe are wrong.
The rollout puts Substack in the same uncomfortable queue as the rest of social publishing: AI text is cheap, detection is probabilistic, and trust is a product feature whether platforms admit it or not. Substack’s answer is to expose a scan option to readers and a process note to writers, while leaving the final interpretation to the humans still doing the reading.
This story draws on original reporting from The Verge.
