Substackers Say New AI Detection Tool Is a ‘Witch Hunt’
404 Media Emanuel Maiberg ● Covered by 4 sources
Substack now scans your posts for AI writing and slaps a percentage score on them. Writers are furious, calling it a witch hunt with shaky accuracy.
Based on reporting by 404 Media, Emanuel Maiberg — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Substack just turned on an AI-detection feature, and a chunk of its writers are not having it. CEO Chris Best framed the move as a defense of "human voices" against a web increasingly clogged with content made by no one. The company partnered with the detection tool Pangram, which now runs quietly in the background and can be triggered by any user to spit out a percentage score estimating how much of a piece was written by a machine.
The reaction from creators has been anything but grateful. Mack Collier, who writes the small newsletter Backstage Pass, said flatly he won't apologize for using AI to sharpen his editing, arguing it makes his writing better, not less his own. Ghostwriter and coach Alice Lemee went further in a video calling detectors "notoriously, wildly inaccurate," warning that a single false accusation can wreck a writer's reputation before they even get a chance to explain themselves.
And that fear isn't hypothetical. Pangram's own CEO, Max Spero, has said the company is still working to shrink its error rate, which it currently estimates at roughly one in 10,000 false positives. That sounds small until you remember Substack hosts a lot more writers than that. Professor Sam Illingworth, who wrote a piece titled "Substack's AI Detector and the Return of the Witch Hunt," put it bluntly: Substack built a machine trained to guess whether a human is hiding inside your prose, then pointed it at everyone's work and asked it to decide who counts as real.
Substack's answer to the backlash is a softer feature sitting right next to the scoring tool: a "How I made this" statement where writers can openly explain their process, AI included. Best insists the company isn't hostile to AI-assisted writing, noting Substack itself uses AI for software and product features, and a company spokesperson stressed that detection doesn't affect discovery and can be turned off or disputed by any writer who thinks a scan got it wrong.
Still, plenty of readers are cheering the crackdown, tired of platforms letting slop pile up unlabeled — the same frustration that's pushed people to build unofficial slop-detection tools for places like Spotify. But the noisy pushback from Substack's own writers is a reminder that slapping a probability score on someone's work isn't the same as solving the problem. Generating AI content at scale is trivial now. Figuring out what to actually do about it, fairly, is still very much unfinished business.
My take — AI-written commentary, not fact-checked reporting
Handing a probabilistic guess a percentage sign doesn't make it a verdict, and platforms keep forgetting that the moment they attach a number to something, people treat it as fact. A false-positive rate that sounds tiny in a press release stops sounding tiny once it lands on a real person's byline and reputation. The "How I made this" disclosure is the smarter move here — it invites trust instead of demanding it — and Substack would be better served leaning into that than pretending a detector built by another AI is some neutral judge of humanity.
Read more about this at: 404 Media