Title: “That Looks Like It Was Written by AI” - So What? Subtitle: The dismissal isn’t an argument. It’s the genetic fallacy in costume. Publication: ELFrederick Publication Date: 2026-04-19 Type: Reactive post Word Count: 1238 Tags: AI writing, logic, dismissal, genetic fallacy, workflow Hero Image: Header.png Thumbnail: Thumbnail.png Related Posts: Three Phases to Publish; Have Some Standards When You Use AI to Write; The Slop Factory Has No Sample Page Status: DRAFT Prep Brief: Prep - That Looks Like It Was Written by AI.md Session ID: 2026-04-19 article-write
“That looks like it was written by AI.”
I have watched that sentence do more work than any argument on the page it was dismissing. It closes threads. It ends comment sections. It lets the reader skip the thinking step and still feel correct. And because the accusation carries the aesthetic weight of a quality complaint, it slides past unchallenged most places it lands.
It shouldn’t. It isn’t an argument. It’s a specific, named logical error, and once you can see the shape, you can’t unsee it.
More often than not, what the accusation really means is that the reader wasn’t equipped to engage the argument.
The Genetic Fallacy
The genetic fallacy is a fallacy of irrelevance in which an argument is dismissed or accepted based on its source rather than its content. Morris Cohen and Ernest Nagel named it in Logic and Scientific Method in 1934. It has sat in the standard taxonomies ever since under the fallacies of relevance, which is the category informal logic invented specifically for moves that look like reasoning and aren’t.
“That looks like it was written by AI” is the pure form. It evaluates the producer. It does not evaluate the produced.
The argument on the page still has a shape. The evidence is still good or bad. The reasoning still holds or doesn’t. None of that has been touched. The dismissal has stepped past the content entirely and rendered a verdict on origin, and rendering a verdict on origin is the one move that has nothing to do with whether the work is any good.
And here is the quiet joke underneath the whole pattern: the “tell” people claim to be spotting is calibrated on the wrong signal. I put the mechanism in a Note back in March:
The model was trained on an enormous corpus of serious published prose, literary fiction and long-form journalism and academic writing, The Atlantic, The New Yorker, the entire universe of careful sentences. Em dashes appear constantly in that register, along with semicolons, balanced clauses, rhetorical precision. The model learned that those features signal sophisticated writing, because in the training data they do. Human raters during reinforcement learning rewarded outputs that sounded literary, and the behavior compounded.
So the detector people think they have is firing on quality signals. They’re spotting variations on the top shelf of the written human record and calling it a machine.
The detector had a window. Default output in 2023 did have recognizable fingerprints, and early readers who noticed them were noticing something real. That window closed. Anyone who has spent a year engineering against those patterns has moved past them, and the people still brandishing the 2023 detector are flagging the writers who did the work and letting the actual slop slide past because it learned to sound human. The intuition is a generation behind the target.
The Parallel
We’ve spent decades teaching people not to dismiss a book because a Black author wrote it, a woman wrote it, a Catholic wrote it. I am not claiming moral equivalence between those bigotries and a category error about where prose came from - the moral weight is not the same and I will not pretend it is. I am claiming the cognitive architecture is identical. The work of that teaching was never about the moral weight of the identities involved. It was about the shape of the move. The move was: decide what the content is worth by checking the label on the producer, and use the label as a substitute for engaging with what was written.
I would want to be judged by the content of my writing, not the nature of its production.
Watch the move in operation. The dismisser encounters a thing. They notice a property of who made it. They rule on the thing based on that property. The content is untouched, because engaging with the content was never the point. The label did the work.
This is why the dismissal slides. The person deploying it doesn’t have to prove the argument is wrong. They don’t have to find the hole in the evidence. They don’t have to notice that they can’t find the hole in the evidence.
They only have to gesture at the label. The label carries the verdict. The thinking step drops out.
And because “AI” isn’t a protected category, the move gets to wear the costume of a quality standard. It’s bigotry with better PR.
What the Process Actually Is
Here is where I am required to grant something, and I grant it freely. Rarely does a response come from someone asking the AI to write something from nothing. If it does, it deserves the ridicule it gets. Slop is real. Slop is boring. Slop deserves every response it receives and I have written pieces that treat it accordingly.
The dismissal I am naming is not the one applied to slop. It is the one applied to work that merits engagement and gets the label instead of a critique.
My own workflow has been under development for about a year now, predating this Substack. It has a research phase, a drafting phase, and a stress-testing phase, and I documented the architecture in a separate piece if you want the operational detail. The part worth naming here is one discipline inside it: I regularly ask AI to write the strongest case against my own position before I write the piece.
That is the single sharpest refutation of the standard objection that AI-assisted work just tells the author what they want to hear. The opposing brief is a built-in adversary. It catches what confirmation bias would hide. It is also, notably, the step the dismisser never imagines exists.
Five hundred fifty-seven thousand words of fiction have moved through this process. Every article on this Substack has moved through this process. The output that reaches the page has been pressure-tested in ways that most unassisted writing never gets, because the assistant is available to argue with and the discipline is built around making it argue. That is not a description of AI writing the piece. It is a description of a writer who has a tireless reader willing to hold the opposing position long enough to find the weak seam in the argument.
The dismissal cannot see any of this, because the dismissal never looked. That is the whole point of the dismissal. It is engineered to not have to look.
The Close
Watch what “that looks like it was written by AI” does when it lands on a piece of writing. It asserts a property about origin. It treats that property as disqualifying. It never engages the argument it claims to be responding to.
The dismisser walked up to a page, saw something they felt equipped to label, deployed the label, and left. The argument on the page is still there. The reasoning is still there. The evidence is still there. None of it has been touched.
This piece included. If the argument above is wrong, it is wrong for a reason you can point to - a premise I got wrong, a definition I misused, a case I failed to account for. The label on the producer is not one of those reasons. It has never been one of those reasons.
You didn’t critique the argument. You couldn’t find one. So you checked the label instead.
You may also like
- Three Phases to Publish - the workflow referenced above, documented in full.
- Have Some Standards When You Use AI to Write - the internal quality gate, before the dismissal ever shows up.
- The Slop Factory Has No Sample Page - when the dismissal lands on work that earns it.