Substack builds native AI detection into its platform, and its CEO has a word for the problem: 'Claudefishing'

Substack's CEO has a name for the problem with AI content: "Claudefishing"

Chris Best, Substack's co-founder and CEO, wants you to picture something uncomfortable. You're reading a post on his platform. The prose is polished, the argument seems original, you're giving it your time and attention. Then you discover nobody actually thought about any of it. A model generated the words, and a person hit publish. Best calls this "Claudefishing" 1.

Substack this week launched an AI detection feature powered by Pangram, an AI detection company. Any reader can now scan posts, notes, replies, and comments longer than 100 words by selecting "Scan for AI text" from the three-dot menu on a post. The tool estimates how much of the content was likely AI-generated or written with AI assistance. It's live on the web and the iOS app, with Android described as coming "soon" 1.

Here's what makes this different from every AI-detection debate that has come before. Substack isn't asking writers to self-police, and it isn't burying the tool in a settings page for authors. It's handing detection directly to the audience. The person reading decides whether to scan. That is a structural choice, and it shifts the power balance in a way that product launches rarely do.

Best frames the problem precisely. "The core problem is not people using AI, or the quality of its output," he writes. "The problem is when there is a mismatch between a reader's expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end. That's Claudefishing" 1.

The term evokes catfishing, and the parallel is unmistakable. The harm isn't the content itself. It's the deception about who's behind it. A reader who enjoys an AI-generated essay wouldn't feel betrayed if they'd known the terms going in. The betrayal lives in the gap between what they assumed, a human mind wrestling with ideas, and what they actually got, a probabilistic text engine optimizing for plausibility.

But Best himself names the limitation, and it's where things get genuinely difficult for thoughtful creators. Pangram can detect whether AI was used to produce text, but it cannot measure whether genuine human care went into shaping that text 1. A writer who drafts with a language model, then edits aggressively, restructures the argument, and owns every idea could trigger the same detection signal as someone who pasted raw model output and hit publish. The tool cannot tell these two people apart. And readers who scan for AI text won't see the distinction either.

That gap puts a specific group in an awkward position: writers who use AI tools as part of a real creative process. They brainstorm with models. They outline with them. They use them to pressure-test structure. Under this system, their work might get flagged, and the flag carries an implicit accusation of thoughtlessness. Substack's answer is a new "How I make this" statement that lets writers disclose their process to readers 1. Writers can also scan their own drafts with Pangram before publishing and report inaccurate results. It's an opt-in disclosure layer sitting alongside an automated detection layer, and the two will inevitably collide.

The broader question is whether competing publishing platforms feel pressure to follow. Medium, Ghost, and WordPress have not publicly announced comparable reader-facing AI detection features. AI detection accuracy remains contested across the industry, with well-documented false positive rates that could alienate the writers who power those platforms. Substack has nonetheless made a notable move: a major publishing platform decided its readers deserve AI transparency built into the reading experience itself. The longer competitors sit out, the more that silence starts to look like a position rather than an oversight.

Best's framing of what's at stake for the platform economy is pointed. Platforms that reward inauthentic content create a downward spiral, he argues, and when readers have to wonder whether what they're reading is real, it undermines trust in authorship and threatens the livelihood of writers, including those who use AI tools thoughtfully to produce work they believe in 1.

The uncomfortable truth underneath all of this is that the tool designed to build trust might also strain it. Writers who use AI thoughtfully now operate under a layer of suspicion they didn't ask for. Writers who never touch AI might get flagged anyway. And readers who scan everything risk treating every post as potentially fraudulent before they've reached the first paragraph. Substack is betting that radical transparency is worth that friction. Whether Pangram's detection is accurate enough to make that bet pay off is a question readers will answer every time they tap that menu.