Table of Contents
Do not index
Can a ghostwriter prove their client's post was actually written by a human, when the platform hosting it has decided a probability score gets to make that call instead? That's the question a growing number of writers are asking after Substack rolled out a built in AI detection score on every post. My answer is no, a detector score cannot prove authorship, and any ghostwriter relying on the platform to vouch for their own work has already lost control of the thing that actually matters most in this business, which is proof.
Substack's tool assigns a probability that a given post was AI generated, and writers, including working ghostwriters, are calling it a witch hunt, because the score can flag human writing as machine made with no way to contest it. A tool that guesses wrong doesn't just embarrass a writer for a day. It attaches a permanent-feeling doubt to work that took real editorial judgment to produce, and that doubt spreads faster than any correction ever does.
Agency operators and ghostwriters are the group most exposed here, because their entire output can get mislabeled by a tool whose maker claims a 1 in 10,000 false positive rate, a number critics say is unproven at scale and known to misfire on non-native English speakers and neurodivergent writers, according to 404 Media's reporting on the backlash. Alice Lemee, a ghostwriter and digital writing coach quoted in that piece, put it bluntly. "These detectors are notoriously, wildly inaccurate. All it takes is one false accusation for a writer to have their reputation almost irreversibly tarnished." That's not a hypothetical risk for anyone billing a client $10,000 or $20,000 a month to protect their voice online. It's the actual downside case.
Why a Detector Score Was Never Going to Settle This
The mistake underneath tools like this is treating authenticity as something you can measure after the fact, from the finished text alone, the same way you'd run a plagiarism check. Writing doesn't work that way. A sentence doesn't carry a visible fingerprint of the three phone calls, the eleven questions, and the two rounds of pushback that went into shaping it. A detector reading only the output has no access to any of that, so it's not actually evaluating authorship, it's pattern matching against training data and guessing. That's a fundamentally different task than the one everyone is using it for, and the gap between those two things is exactly where writers with legitimate human process get caught.
I don't ask my clients or my team to trust a score, because I don't trust one myself. What I call the Origin Trail is the alternative: every piece of client content carries a documented path back to its source, the voice notes, the call transcript, the specific detail the client mentioned once that only they would have said. None of that lives in the published post. It lives in the process behind it, and that's the part a detector can never see, because it was never designed to look there. When a client or a platform questions a piece, the Origin Trail is what settles it, not a probability score generated after the fact by a tool with an unverified accuracy claim.
Building an Origin Trail Into a Ghostwriting Retainer
This matters most for ghostwriters and small agencies managing five to fifteen client retainers who publish under someone else's byline, because volume is exactly what makes memory an unreliable defense. You cannot recall, six weeks later, which specific call produced which specific line across fifteen different voices. The Origin Trail has to be built into the retainer from day one: source material archived per post, drafts saved with timestamps, client call notes tied directly to the pieces they informed. It's closer to how founders should position on LinkedIn as practitioners first, with specifics that only someone doing the actual work would know, rather than as generic thought leaders trading in phrases anyone could produce. Specificity is what makes writing defensible, and specificity is exactly what an Origin Trail preserves and a detector score cannot detect.
This isn't for writers who don't keep any raw material from their client calls or interviews, because there's nothing to point to if a challenge ever comes. It also isn't relevant for people writing purely for themselves with no client relationship or reputation exposure riding on the outcome. If your only defense against an accusation is trust me, you were already exposed before Substack built a button for it.
The platforms are going to keep shipping detection tools, because AI generated content at scale is a real problem for them and they need something to point to. None of those tools are going to get meaningfully more accurate anytime soon, because the underlying task, guessing authorship from text alone, is harder than it looks and easier to get wrong than any vendor wants to admit. Ghostwriters who build proof into their process now, instead of waiting to see how good the detectors get, are the ones whose retainers survive the next false flag. Everyone else is betting their reputation on a coin flip they don't control.
