I want to document what happened to me last year, because I think it illustrates something that’s going to become much more common.
I submitted a manuscript to a mid-tier linguistics journal. Eight months in review. The final email wasn’t a rejection - it was a request to “address concerns about the origin of the text,” along with a screenshot of an AI detector score. 73% AI probability. On a paper I wrote entirely myself over fourteen months.
My writing is dense. Highly structured. I use parallel construction because it mirrors the corpus analysis methods I’m describing. I eliminate hedges in prose because the hedges belong in the methodology section. According to AI detectors, this sounds like a language model.
I appealed. I submitted drafts, annotated notes, a writing log I keep out of habit, and a statement from my co-author. The journal’s response was to send the appeal to a second detector, which returned 68%. They declined to publish.
Here’s the part that stays with me: I couldn’t prove I wrote it. Not definitively. The evidence I had was circumstantial. Drafts can be fabricated. Writing logs can be fabricated. My co-author can be wrong or lying. There is no handwriting sample for prose. No biometric equivalent. And I spent fourteen months on that work.
The epistemic problem is deeper than the technology problem. Even if detection tools improve, the verification question remains. How do you prove authorship of text? We’ve never had to answer that question before. Text was assumed to originate from whoever signed it. That assumption is now broken and I’m not sure what replaces it.
I’m curious whether anyone else has hit this in academic or professional contexts. And whether there’s a reasonable way to build a verification practice before you submit, rather than scrambling after the fact.
This is the case that haunts me. I sit on an Academic Integrity Committee and we have the inverse problem: students saying they wrote something that the detector says they didn’t. And we have the same evidential problem from the other side.
A detector score is not proof. We decided that early. But it’s also not nothing. So we investigate, which means asking for drafts, notes, a verbal walkthrough of the argument. Some students can do that and some can’t. And you can’t always tell from the walkthrough whether the difficulty is guilt or just a student who wrote something and didn’t fully understand it.
Your situation is worse because you understand your work completely and still lost. That’s the failure mode I worry about most. Not the cheater who gets away with it. The honest person who can’t prove it.
As a PhD candidate this genuinely frightens me in a way that’s hard to explain to people outside academia.
I use AI for literature organization and outlining. I write the actual text myself. But I’ve tested my own writing through detectors and gotten scores in the 60s. My advisor writes in a way that I’ve absorbed over four years. She writes formally. I write formally. Apparently that reads as artificial now.
The thing you said about the assumption being broken - that’s exactly it. Authorship was always somewhat constructed but we had a consensus fiction that text and name were connected. Now we don’t. And the systems we’re building to replace that consensus are doing real harm to real people’s careers before they’ve been validated at all.
Coming at this from content work rather than academia, but same underlying problem.
Clients run detection on my deliverables. Some of them have told me upfront, some I’ve found out later. I pass consistently, but only because I’ve built a workflow that accounts for it. Not because I can somehow prove I’m the author.
The idea that you can prove authorship is kind of a fantasy at this point. What you can do is manage the probability. What your journal did was treat a probability estimate as a verdict. That’s the actual problem and it’s not a technology problem, it’s a policy problem. Detectors were never designed to be arbiters of authorship disputes. Someone decided to use them that way.
The fiction writing community is dealing with a version of this too. Literary magazines that have added AI submission bans are running detectors on everything. I know writers who’ve been rejected or blacklisted who didn’t use AI at all. Clean, distinctive voices that apparently pattern-match to something the models produce.
The trust damage runs in both directions. Readers and editors can’t verify authors. Authors can’t verify their own originality anymore, in the sense that originality is now entangled with detectable patterns rather than actual process.
I don’t have a solution to offer. I just think the “just don’t use AI and you’ll be fine” logic is badly oversimplified.
What you went through is a serious institutional failure and I want to be direct about that. A 14-month research project being declined on the basis of a detector score that the journal itself couldn’t validate is not defensible editorial practice.
The publishing world is heading toward this too. I’ve already had authors ask me to sign something attesting that the manuscript was human-written before I’d even looked at it. They want cover. They’re scared. Everybody is scared and nobody has built the infrastructure to handle it responsibly yet.
The verification question you’re raising doesn’t have a clean answer. Timestamped draft commits, version histories, real-time writing sessions - none of those are tamper-proof either. At some level this is going to require trust again, but rebuilt on better institutional frameworks rather than detector scores.