Substack introduces AI detection to combat ‘Claudefishing’
▼ Summary
– Substack launched an AI-detection feature called “Claudefishing,” powered by Pangram, to let readers scan posts for AI-generated text.
– The tool estimates human versus AI contribution for text over 100 words published from today, showing results only to the requesting user.
– Substack’s CEO clarified the problem is not AI use but readers unknowingly consuming content with no human thought behind it.
– AI detection is not perfect; critics warn it can be evaded by fine-tuning AI and may lead to false accusations.
– Substack bets that proving human authorship becomes a valuable, scarce commodity in an AI-saturated market.
Substack now has a name for the sinking feeling when a beautifully written post turns out to be machine-made: Claudefishing. And the platform is giving readers a direct way to catch it, rolling out a detection tool powered by the AI scanner Pangram.
The newsletter service has introduced a new AI-detection feature. In a blunt post titled “Against Claudefishing,” CEO Chris Best explained that readers can now scan posts, notes, replies, and comments. The tool provides an estimate of how much of the text was written by a human and how much was generated with AI assistance.
The scan works on any text longer than 100 words that has been published from today onward. The results are visible only to the person who requested the scan. The feature is live on the web and iOS, with Android support coming soon.
What Claudefishing means
Best is careful to clarify that the issue is not AI itself. Substack uses AI for building software and product features, he noted, and many writers use it thoughtfully. The real problem is the disconnect: readers invest attention in content that contains no genuine human thought. “Claude has a lot to offer,” he wrote. “But when I want Claude’s opinion, I’ll ask Claude.”
The stakes are rising fast. Pangram estimates that as much as 40% of text on some social platforms is now AI-generated. Substack, which markets itself as an economic engine for culture built on trust between people, wants to prevent its feed from turning into that kind of slop.
Not a ban
The tools work both ways. Writers can run the detector on their own drafts before publishing, add a “How I make this” statement to explain their process, and report any scan of their work they believe is inaccurate. Substack frames the entire effort around a simple principle: people should know what they are getting.
It is an unusually direct move in a market saturated with AI-generated feeds. The post even took a swipe at LinkedIn, the platform many see as ground zero for machine-written earnestness.
The catch
The problem is that AI detection is not settled science. Pangram is accurate, but not perfect. The Atlantic has warned that over-trusting such tools risks turning accusations into a witch hunt. Substack itself concedes the tool cannot determine whether real care went into a piece, only whether AI shaped the words.
Critics go further. Any detector, developer Perry Metzger argued, lets you fine-tune an AI to evade it, so such tools cannot work for long. Researcher Mor Naaman called detection a losing battle: if readers scan, writers will scan too and route around it. As one Substack commenter put it, the presence of AI does not prove the absence of a human.
A very human business
Underneath it all sits a bet. As Business Insider’s Peter Kafka put it, if AI wins, “made by humans” could be a very good business. Substack is wagering that proof a person did the work becomes the scarce, sellable thing. That the fight is playing out over newsletters, of all places, only shows how far the question now reaches.
(Source: The Next Web)
