Substack now scans posts for AI fingerprints — 100 characters is all it takes

Substack just made a fascinating bet against its own short-term interests. The company is integrating Pangram, an AI detection engine, directly into its app. Any post, reply, or comment longer than 100 characters will now carry a readout estimating how much of the text came from a human versus a machine. It’s a tool that could, at least initially, embarrass the platform by exposing just how many newsletters are running on autopilot.

Chris Best, Substack’s CEO, framed the move as a return to the company’s original pitch. In a chat with Pangram founder Max Spero, he said the software should handle everything except “the hard part” — the idea worth reading. Detection, in this view, isn’t punishment. It’s a spotlight. Substack is also giving writers the option to attach an AI author’s note to their work, a disclosure mechanism that puts the onus on creators to explain their process rather than hide it.

The timing is no accident. Across media, the pendulum is swinging toward labeling. Social platforms tag AI-generated images and video. Spotify is penalizing AI music. Substack’s move is the textual equivalent, and it arrives just as “AI slop” becomes a term of genuine reader fatigue. The risk is real: if subscribers see that a favorite newsletter scores high on the machine meter, trust evaporates. But the alternative — a platform slowly flooded with synthetic prose — is far worse for a business built on the promise of independent human voices.

Publishers do get some control. They can pre-scan their own drafts using Pangram and even flag and remove scans they believe are wrong. It’s a tacit admission that detection isn’t perfect. But the message to readers is clear: Substack is willing to make the invisible visible, even if it stings a little. The feature is rolling out in the app now.