Substack is not banning AI writing. Its new move is narrower and more telling: readers can scan a post, note, reply or comment and see an estimate of how much text may have been written or assisted by AI.
The feature launched on July 21 with AI-detection company Pangram. It does not decide whether a post has value. It gives readers a signal about machine involvement. For a subscription platform built on author-reader trust, that distinction matters.
The button starts with readers
Substack co-founder and CEO Chris Best calls the mismatch “Claudefishing.” In his words:
The core problem is not people using AI, or the quality of its output. The problem is when there is a mismatch between a reader’s expectation and reality.
The scan applies to posts, notes, replies and comments over 100 words, published from July 21 onward, and the result appears only when a reader asks for it. The Verge reports that the feature is live on web and iOS, with Android coming later.
Substack is also adding a “How I make this” statement so creators can explain whether they use AI for research, editing, translation or not at all.
Platforms are competing over human provenance
Business Insider notes that creator platforms are taking different approaches to AI slop. YouTube is clarifying its inauthentic-content policy, TikTok requires AI labels and is testing spam detection, Pinterest lets users limit generative AI in feeds, and Meta applies AI info labels to some content.
Substack’s risk is different. Its business depends on readers believing that a particular person’s judgment is present. AI can help a writer work faster, but hiding the human role makes the subscription relationship fragile.
The detector is not a judge
Pangram says its detector reaches 99.98% accuracy and has been independently verified. Substack uses more careful language: Pangram can detect likely AI use, but it cannot measure human care, authorship quality or whether AI was merely a source or editing aid.
That limitation is central. A scan score should not become a moral grade. Substack therefore lets creators scan drafts, report mistakes and remove scans they believe are wrong.
A new label market
The larger signal is commercial. “Made by humans” may become a selling point for some newsletters, communities and brands. Substack is shifting the question from whether AI is allowed to how much provenance readers want.
The next checks are concrete: whether scan results affect recommendations, whether creator process statements become common, whether readers get feed-level AI controls, and whether appeals work when detectors miss.
AI writing will not recede because of one button. But Substack has put the authorship question back in front of readers: who is speaking, how much human judgment is present, and whether that is what the reader came to pay for.
Sources: Substack official blog, The Verge, Business Insider, Pangram official information, CocoLoop; checked launch scope, 100-word scan threshold, Pangram partnership, creator disclosures, platform AI-content controls and detector limitations.