Substack Now Lets Readers Scan Posts for AI-Written Text
A new detection tool built into Substack will estimate how much of any post was written by AI, giving readers a way to check before they invest their time.
Key points
- Substack began rolling out an AI-text detection tool on its web platform and iOS app on Tuesday.
- The tool is powered by Pangram, an AI detection company, and works on any content longer than 100 words.
- Writers can scan their own drafts with Pangram before publishing and flag results they believe are wrong.
- Substack CEO Chris Best coined the term "Claudefishing" to describe content that tricks readers into thinking a human wrote it.
Substack wants its readers to know whether a real person wrote what they are reading, or whether a machine did most of the work.
The newsletter platform has started rolling out a built-in AI detection tool, powered by a company called Pangram, which specialises in identifying AI-generated writing. Readers can trigger a scan on any post, note or reply longer than 100 words. Tap the three-dot menu in the top-right corner of a post and choose "Scan for AI text." The tool returns an estimate of how much writing could have come from an AI, or been heavily assisted by one. Android support is coming soon.
Substack co-founder and CEO Chris Best framed the move as a trust problem, not a technology ban. "The core problem is not people using AI, or the quality of its output," Best wrote in a blog post. "The problem is when there is a mismatch between a reader's expectation and reality."
Best gave that mismatch a name: Claudefishing. It's a play on catfishing, where someone pretends online to be someone they're not, except here a writer pretends human thought went into text that a large language model (the technology behind chatbots like ChatGPT) actually produced.
The tool has real limits. Pangram can tell you whether AI shaped the words on the page. It can't tell you whether the writer spent hours directing and editing that output carefully. A thoughtfully assisted piece and a lazily generated one could score similarly. Best is open about that gap.
To handle it, Substack is adding a voluntary "How I make this" statement where writers explain their process. Think of it as a nutrition label for creative work. We covered the limits of AI detection from a different angle when an Australian company certified the Pope's writings as genuinely human-authored in July, and the awkwardness of that story hasn't gone away.
Should readers trust the scan results?
Treat them as a useful signal, not a verdict. Pangram's estimate is a probability. Best himself notes it can produce inaccurate results, which is exactly why writers get an option to flag a score they think is wrong.
For paying subscribers or anyone giving a writer their attention, there's now a quick way to sense-check whether a human crafted what they're reading. That's more than most platforms offer. Instagram, for instance, still won't let you filter AI posts out of your feed.
Best put it plainly: "Platforms that reward fakeness will create a race to the bottom."
The practical question to watch isn't whether detection works perfectly. It's whether writers start gaming the score the moment they learn how Pangram thinks.
