What Substack's percentage says, and doesn't
Substack partnered with Pangram and added a tool that scans posts, replies and comments longer than 100 words published from July 21 onward, returning an estimate of how much of the text is human versus AI-assisted. The company's position is clear: “people should know what they're getting.” But Substack itself acknowledges Pangram's limits, echoing criticism that such tools are far from perfect and can't guarantee an accurate result.[1]
My skepticism is an occupational deformation, I know. But a percentage is not a source. Telling a reader “30% of this is AI” doesn't answer which model, whose data it was trained on, or under what consent — it just puts a label on the output. A label can be useful, but it isn't a paper trail.[1]
Deezer's half: whose voice, whose payment?
Deezer's own detection technology found more than 90,000 AI-generated tracks uploaded daily in June 2026 — over half of all daily uploads, up sharply from 10% (10,000 tracks a day) in January 2025. CEO Alexis Lanternier said: “Now that half of all daily uploads are AI-generated tracks, we are taking additional steps to safeguard the rights of artists and songwriters.”[2]
In my July 21 column I wrote that a court ruling that pays out and a platform policy that withholds pay were both solving the source-tracing problem in the currency of money, not in the currency of an actual answer — and that we might never fully learn where the exhaust goes. Substack and Deezer are now trying a third route: detect and label. Both are useful; neither names the source. My next test: will Deezer's “additional steps” turn into a real payment or attribution mechanism for artists, or just another gatekeeper that cuts off money, the way YouTube's rule did?[1], [2], [3]