Substack launched an AI content detection feature on July 21, powered by Pangram, that allows readers to scan any post, note, reply, or comment published on or after that date and receive a percentage estimate of how much of the text is human-written versus AI-generated or AI-assisted. The feature is available through the three-dot menu in the Substack Reader on web and iOS, with Android support coming later. It applies to text longer than 100 words. CEO Chris Best introduced the launch in a post titled Against Claudefishing – his term for what he described as the gap between a reader’s assumption that a human wrote something and the reality that no one did. NEWSCENTRAL reads the Substack move as the first major publication platform to take a formal institutional position on AI content authenticity in the direct writer-reader relationship that distinguishes the newsletter economy from traditional media, and the reaction from creators – ranging from furious to cautiously supportive – as evidence that the platform has correctly identified a genuine trust problem, even if the tool it has introduced to address it is imperfect.
The commercial logic behind Best’s framing is cogent and reflects the specific economics of the paid newsletter model. Substack’s revenue and its writers’ revenue both depend on subscribers paying recurring fees for access to a specific human perspective. That model collapses if readers cannot distinguish between content that reflects the genuine judgment, experience, and voice of the person they subscribed to and content that is an AI-generated approximation of that voice. Best articulated the distinction with unusual clarity: software should do everything except the hard part, and the hard part – the idea that is worth reading, worth caring about, worth sharing – is precisely what readers are paying for. A platform that allows that hard part to be systematically replaced by AI without disclosure is, in his framing, defrauding its own users.
The creator response has been sharply negative for a specific and legitimate reason. AI detection tools are not reliable enough to function as authoritative judgment in high-stakes reputation contexts. Pangram, the detection partner Substack has chosen, uses a classifier neural network trained on pre-2021 human text – a training cutoff that means the detector may not recognize newer AI writing patterns, and may falsely flag human writers whose style happens to be precise, structured, or stylistically similar to AI output. Substack has tried to address this concern by giving creators the ability to run the detector on their own drafts before publishing, contest misclassifications, and disable the AI detection feature for readers entirely. Authors can also add an optional How I make this statement disclosing their writing process. Nathan Clark, Enterprise IT and Systems Architecture Analyst at NEWSCENTRAL, observes that the ability to disable detection creates an immediate adverse selection dynamic: the writers most likely to use that option are precisely those whose content would not survive the scan, which means readers who notice that a publication has disabled AI detection will draw the inference that the tool’s absence is not value-neutral.
The broader platform context that Substack is navigating makes the timing of this launch commercially sensible even if it generates near-term creator friction. LinkedIn has begun reducing the reach of AI-generated posts. Meta is testing its own detection tool. YouTube is restricting monetization of AI-generated content. The emerging platform norm that AI content is permitted but must carry a label is consolidating across the major creator platforms, and Substack’s choice to adopt this norm before it is externally mandated – and to give creators agency in how they participate – positions the platform as a leader rather than a reactive follower. The alternative – allowing the paid newsletter ecosystem to fill with undisclosed AI content until the resulting subscriber backlash forces a response – would be more damaging to the Substack model than the creator friction the detection launch is generating.
NEWSCENTRAL considers the adverse selection dynamic embedded in the opt-out feature the most commercially interesting design tension in the Substack AI detection launch: a tool whose most commercially motivated users can disable it is a tool whose results carry different information content depending on whether the author has left it enabled or turned it off, and readers who understand that dynamic will quickly learn to treat opt-out as a signal in its own right.
The naming choice – Claudefishing rather than AI-washing or some more neutral formulation – is analytically interesting in what it reveals about how Best conceptualizes the problem. Claudefishing implies that the deception being addressed is specifically the use of sophisticated frontier AI to simulate a human voice convincingly enough to extract a recurring subscription payment, rather than the general use of AI tools to assist with writing. NEWS CENTRAL considers that framing commercially accurate as a description of the most commercially damaging form of AI content abuse in the paid newsletter context, and a recognition that the proliferation of capable AI writing tools has made the deception economically accessible to any creator who chooses to pursue it.