Substack partners with Pangram to fight AI slop with content detection — Chris Best explains the Claude-fishing problem

Substack has partnered with Pangram Labs to add AI content detection to its platform, letting readers scan posts, replies, and comments for an estimate of how much was written by a human versus AI. Writers can also attach a "how I make this" disclosure statement to their work. Chris Best, Substack's founder, frames the feature as a transparency tool rather than a prohibition.

The problem Best calls "Claude-fishing"

The motivation isn't AI use itself — it's undisclosed AI use. Best describes the experience of reading something online, realizing mid-scroll that it was entirely AI-generated, and feeling the connection to the supposed author dissolve. That moment of deception is what he calls Claude-fishing, and he sees it as the core damage to reader trust, not AI writing per se.

Pangram's own research, drawn from its browser extension, found that roughly 40% of long-form text on LinkedIn is AI-generated. Substack came in around 10%. Best says Substack's number isn't alarming yet, but the partnership is designed to keep the bar high before the problem compounds.

“The guys at Pangram did this study and they found that on some platforms like LinkedIn, 40% of the long form text is generated by AI. The problem is not people using AI. The problem is when you have the deception that comes with expecting that I've got a real connection to a person and I'm actually reading something that's written by no one. I think of this as clod fishing.”

What counts as slop

Best draws a sharp line. Slop isn't content made with AI; it's content nobody believes in — the "generate 1,000 viral posts, make no mistakes" approach where there's no human perspective behind the output. Writers using AI as a research partner, editing layer, or production tool, while being transparent about it, don't fall into that category in his view. He points to writers on Substack whose first language isn't English, or who are doing deep research and generating final prose, as legitimate examples. The test he offers is simple: if full disclosure wouldn't ruin it for the reader, the AI use is probably fine.

The underlying thesis

Best argues the value of a trusted human perspective goes up, not down, as AI-generated content becomes abundant. If a reader can get any analysis from a model directly, what they're paying for on Substack is the specific person — their judgment, their curation, their point of view. He uses Tyler Cowen as an example: a subscription to Cowen's prompts and research process would be valuable precisely because readers are buying access to his thinking, not just the output.

On the business, Best says Substack is growing. He doesn't give specific figures but pushes back on the narrative that AI erodes demand for newsletters, arguing the opposite dynamic is more likely as undifferentiated content floods other platforms.

Norms still forming

Best acknowledges there's no settled consensus on where the line sits — between deep-research AI-assisted work and pure automation, between a personal essay and a breaking-news roundup. His framing is that new norms are being worked out in real time, and Substack's role is to give readers the information to make their own judgments rather than to impose rules. The Pangram feature is the first step in that direction.