The Substack platform has launched transparency tools to identify artificial intelligence-generated content as part of a fight against the "Claudefishing" phenomenon—the mass generation of text designed to mimic human involvement.
What Happened
In partnership with the service Pangram, Substack has implemented a system to scan notes, replies, and posts longer than 100 words to estimate the proportion of AI-generated text. The functionality is opt-in: authors can add explanations regarding their content creation process via a "How I make this" section and can contest erroneous detection results.
Context
The "Claudefishing" problem represents a direct attack on the trust economy of platforms, where fully AI-generated content is presented as original authorship. The implementation of such tools is an attempt to create an authorship verification layer and protect creator revenue from AI noise.
Why It Matters for the Industry
This move sets a trend for content platforms to implement verification mechanisms and distinguish "pure" human content from hybrid AI-assisted content. The use of Pangram instead of in-house development indicates a growing demand for specialized APIs for AI content detection.
Why It Matters for Users
Readers gain the ability to see AI-generation labels, which helps them more effectively filter high-quality analysis from mass machine spam and better understand the nature of the content they consume.
What Is Not Yet Known / Limitations
A fundamental contradiction exists: while developers see this as creating new "Proof of Human" UX patterns, ML experts emphasize the technical impossibility of guaranteed detection given the constant evolution of LLMs.
Sources
Author
Look at AI, Editorial Team