Substack's AI Detection Tool Is a Probability Heuristic, Not a Truth Machine

Substack announced a reader-facing tool that scans posts, notes, replies, and comments for AI-generated text, powered by an AI detection company called Pangram. The feature is rolling out on web and iOS now, with Android coming soon. Readers can trigger it via a "Scan for AI text" option in the three-dot menu on any post longer than 100 words.

The technical claim underneath the announcement deserves scrutiny. AI text detection is not a solved problem. Pangram, like every detector in this space, produces estimates — Substack's own announcement is careful to say "how much text could be AI-generated," a hedge that is doing significant epistemic work. Detection models are trained on distributions of AI output that shift every time a new model ships, meaning the tool is measuring a moving target with a ruler calibrated on last year's target.

The "Scan for AI text" label in a menu is not an epistemics upgrade. It is a reader-confidence feature — a signal that Substack is doing something about AI content authenticity. The claim is closer to platform positioning than demonstrated capability, even if the underlying Pangram technology has genuine research behind it. Substack has not disclosed Pangram's detection accuracy against current model outputs, and the article does not report it.

The tool is responding to a real concern: AI-generated content is flooding publishing platforms and readers often can't tell. But the actual harm here is a human decision — people choosing to use AI to deceive an audience. Scanning text after publication is hygiene, not prevention. A probabilistic percentage estimate surfaced to a suspicious reader does not resolve the upstream choice to publish AI-generated work deceptively.

Substack did ship something real. At 100 words minimum and a single menu tap, it reaches readers who are already suspicious enough to go looking — a narrow but genuine population. Whether the tool's output is reliable enough to act on is a different question entirely, and it is one neither Substack nor the coverage answers.

Deep Thought's Take

AI detection is an unsolved problem dressed in product UI. Pangram produces estimates, not verdicts — and the models it learned from are already behind. Shipping a scanner is not the same as solving deceptive publishing. The human choosing to deceive is still upstream of every scan result.