Substack added an AI detector, and its writers are not having it

Substack rolled out an AI-detection feature on 21 July, built with a company called Pangram. Any post, note or comment longer than 100 words can be scanned, and the tool returns a percentage-style score for how likely the text is to be AI-generated. The result is shown to readers only when they ask for it, the feature applies only to material published after 21 July, and writers can switch it off for their own work. None of that has calmed the reaction. Writers across the platform have called it a witch hunt, and the objection is not mainly philosophical. It is about accuracy.

The evidence they point to is uncomfortable for anyone selling detection. A widely cited 2023 Stanford study ran seven AI detectors against essays written by humans whose first language was not English, and on average 61 percent of those genuine human essays were flagged as AI-written. The failure mode is not random. Detectors key on things like predictable vocabulary and low sentence variation, which describes AI output but also describes a careful writer working in a second language, or anyone writing plainly on purpose. The asymmetry of consequences does the rest: a missed detection costs nothing, while one false accusation can follow a writer around permanently, and there is no clean way to prove a negative.

The deeper problem is that AI detection is not a solvable engineering task in the way spam filtering is. There is no watermark in ordinary text, so a detector is really a style classifier making a probabilistic guess, and it is guessing about a moving target that gets more human-sounding with every model release. It also cannot see the distinction that most readers actually care about. Text drafted by a person and tidied up by a model, and text generated wholesale from a one-line prompt, look far more similar to a classifier than they do to a reader. A single percentage score flattens all of that into a number that feels authoritative and is not.

What this means for you: if you write anything in public, assume your work will get scanned by something at some point, and that the score can be wrong in your disfavour. The practical defences are boring and effective: keep drafts, version history and notes, so you can show your process rather than argue about a number. Do not rewrite your natural voice to please a classifier, which is the trap several Substack writers described. If you run a publication, a school or a team, the sensible policy is to treat detector output as a weak signal that might start a conversation, never as evidence that ends one. And if you use AI in your writing, the thing that actually protects you is saying so.

Sources

• Substackers Say New AI Detection Tool Is a ‘Witch Hunt’ (404 Media)
• Substack adds AI detection tools, but creators can disable them (Jane Friedman)
• GPT detectors are biased against non-native English writers (Stanford, arXiv)