Detection tools have existed for years. What changed is who holds them.
Substack has integrated Pangram into its app, and the scanner is pointed at the reader rather than at an editor. Anyone can run it against a post, a note or a comment over 100 characters and get back a percentage estimate of how much of the text was AI-written. It applies to content published from 21 July 2026 onward.
That is a different thing from a newsroom checking submissions. The audience now has the button.
The accuracy problem is real and cuts both ways
Pangram says its current model was trained on human text written before 2021 and reports a 0.01% false positive rate. That number deserves scepticism, and not only from people who want it to be wrong: independent studies of comparable detectors have repeatedly found that advertised false-positive rates understate what happens on real-world text. Critics of the Substack rollout have raised exactly this.
So two things are true at once. Detection is now visible to your readers, and detection is not reliable enough to be treated as proof. A flag is evidence of a statistical resemblance, not a finding of fact.
The practical consequence is unpleasant for honest writers: if you write in a clean, structured, unadorned style — which is what most business writing advice tells you to do — you look more like a model than someone who rambles. Non-native English writers get flagged disproportionately for the same reason. You can be entirely human and still light up the scanner.
The three forces arriving together
Taken alone, one platform feature is not a trend. It is not alone.
Platforms are adding reader-facing detection. Substack is the notable one so far, and features like this rarely stay confined to one platform.
Regulation now requires disclosure in some contexts. The EU AI Act's Article 50 transparency obligations took effect on 2 August 2026, including a requirement to disclose AI-generated text published to inform the public on matters of public interest, unless a human took editorial responsibility for it.
Search rewards effort and punishes volume without it. Google's August spam update demoted high-volume, low-originality content while AI-assisted pages carrying genuine original material held or grew.
None of these bans writing with AI. All three penalise pretending you did not.
The move that actually works
Substack paired the scanner with an optional AI author's note and a "How I make this" statement, letting writers describe their process up front. Copying that pattern on your own site is the single highest-value response available, and it costs an afternoon.
The reasoning is straightforward. A detector flagging a post you already described as AI-assisted tells your reader nothing they did not know, and the flag carries no weight. A detector flagging a post that implied a human wrote every word costs you trust you cannot buy back. Disclosure converts a potential scandal into a footnote.
It also happens to be the honest option, which is worth something on its own.
What to write on that page
Vague statements do not help. "We use AI responsibly" is worse than nothing, because it reads as a hedge. Specificity is the whole point.
Say which parts are machine-assisted — research, drafting, editing, all three. Say who checks the facts and what happens when a claim cannot be verified. Say whether a named human takes editorial responsibility for what publishes, since that is precisely the condition Article 50 turns on. And cite your sources on the posts themselves, because a stated, checkable source is the one signal a detector cannot manufacture and a reader can verify in a click.
If you are flagged and you did write it
Keep a paper trail as a matter of routine. Drafts in version control or a document with edit history are ordinary business practice and happen to be the only rebuttal that works. Arguing with a percentage is unwinnable; showing the draft history is not.
And resist the temptation to write badly on purpose to defeat a classifier. Deliberately degrading your prose to look human is a losing trade against readers, who are the actual audience.
The bottom line
The question is no longer whether you can publish with AI without anyone noticing. Assume a reader can check, assume the checker is imperfect in both directions, and make the answer boring by saying what you do before anyone asks. Disclosure plus visible sources is cheap, survives a false positive, and satisfies the regulation as a side effect.