Substack implements AI detection to combat 'Claudefishing'.
Substack has coined a term for the moment you realize that a post you appreciated for its genuine voice was actually generated by a bot: Claudefishing. The platform is now providing readers with a button to detect this, enabled by the AI detector Pangram.
They have introduced a feature for detecting AI-generated content. In a pointed message titled “Against Claudefishing,” CEO Chris Best stated that readers can now scan posts, notes, replies, and comments. This tool assesses the proportion of text written by a human versus that which was generated with AI assistance.
The detection functions on text longer than 100 words, published from today onward, and the results are visible only to the requesting user. It’s currently available on the web and iOS, with Android support to come.
Understanding Claudefishing
Best emphasizes that the issue is not with AI itself. He noted that Substack utilizes AI to develop software and product features, and numerous writers thoughtfully incorporate it into their work. The challenge lies in the disparity: readers focusing their attention on content that lacks human input. “Claude has a lot to offer,” he mentioned. “But when I want Claude’s opinion, I’ll ask Claude.”
The stakes are rising rapidly. Pangram estimates that up to 40% of text on certain social platforms is now generated by AI. Substack, which positions itself as a cultural economic engine rooted in trust, aims to prevent its feed from devolving into such content.
Not a prohibition
The tools are beneficial for both writers and readers. Authors can use the detector on their drafts prior to publishing, include a "How I made this" statement to detail their creative process, and raise concerns about any scans they believe are inaccurate. Substack emphasizes a straightforward principle: people deserve to know what they are engaging with.
This move is notably direct in a market oversaturated with AI-generated content. The announcement even criticized LinkedIn, a platform many regard as a hub for machine-produced sincerity.
The challenge
However, AI detection is not an exact science. While Pangram provides reliable results, it is not infallible. The Atlantic has cautioned that over-reliance on these tools could lead to a witch hunt mentality. Substack acknowledges that the tool cannot ascertain whether genuine effort was put into a piece, only if AI influenced its wording.
Critics are even more skeptical. Developer Perry Metzger argued that any detection system allows for AI to be fine-tuned to bypass it, meaning such tools may not be sustainable in the long run. Researcher Mor Naaman suggested that detection is a futile endeavor: if readers choose to scan, writers will do the same and find ways around it. As one commenter on Substack noted, the presence of AI does not necessarily indicate the absence of human effort.
A fundamentally human operation
At its core, this represents a gamble. As Business Insider’s Peter Kafka observed, if AI prevails, “made by humans” could become a highly valuable offering. Substack is betting that evidence of human authorship will become a rare, marketable commodity. The fact that this debate is occurring in the realm of newsletters highlights the extensive implications of this issue.
Other articles
Substack implements AI detection to combat 'Claudefishing'.
Substack has teamed up with the detector Pangram, allowing readers to check posts for the proportion of content produced by humans versus AI. This form of deception is referred to as Claudefishing.
