On the difference between a pattern and a fraud, and why a tool that can’t tell them apart has been handed the authority to decide which one you are.
This week I ran three of my own articles through an AI detector. Nerd that I am, I got curious about a tool I could trial and see whether it was any good. One article I hadn’t published yet, two had been live for months. One of those was published since before large language models could write a coherent paragraph, the other was one where I had used a proofreading tool that comes with my preferred authoring app.
After adding the pieces and running a scan, I expected the flagged sections to track meaning: the parts where I overreached, where a metaphor got away from me, where I said something I didn’t fully earn. Instead, I watched sections light up in almost exact proportion to how tightly organised they were. Cut one self-important sentence from a loosely built piece and the score dropped by more than twenty points. Cut the same kind of sentence from a piece built as a mechanism mapped across three deliberate scales, and the score remained the same. Because there was no loose, unstructured stretch of text for the cut to register against, it simply didn’t influence the outcome. Content had almost nothing to do with it. Architecture did.
That’s a strange thing to happen, because architecture is not the same thing as fraud. But increasingly, it seems, it’s being read as its evidence.
Ask anyone who distrusts AI-assisted writing what gives it away, and the em-dash comes up before anything else does. I’ve never used them often, and I don’t especially like them unless a sentence genuinely needs the interruption. I haven’t stopped using them because of what people now seem to assume about a writer who does. But I find it strange, and a little telling, that a punctuation mark can become a loyalty test: not “is this good“, but “did you self-censor enough to prove you’re one of us“. The mark didn’t change; what changed is that something else started using it convincingly, and the suspicion ran backward from the imitator to the tradition it learned from.
Repetition, parallel structure, the rule of three, the aphoristic closer, the long sentence that turns on a colon: these are not new inventions. They are old tools; some of them centuries old, refined by people who cared about the rhythm of a sentence long before anyone needed to prove they were human to be believed. Large language models are fluent in them because they were trained on a great deal of writing that used them well. The tools got good at the craft. Now the craft is treated as evidence of the tool. The writers who reach for it honestly, because it says what they mean more precisely than the flatter alternative, get asked to prove a negative.
I was taught, when studying art and design, that form follows function. That the shape of a thing should be explained by what it does. But there’s a second half to that idea, one that matters more here: function follows form when aesthetics and emotion enter the equation. When the shape itself is the point, that call belongs to the person making the thing, not the person judging it afterward. If it worked the other way – if the critic decided what a thing was allowed to look like based on what most people found palatable – we wouldn’t be talking about craft anymore. We’d be talking about taste. And taste, put to a popular vote, tends toward the lowest common denominator. If the masses decided what food we were permitted to make, we’d all be eating worse, more often, and calling the alternative elitist.
That’s the part of this that actually interests me, maybe even more than the detector itself. A detector being unreliable is a technical problem, and probably a temporary one (it even scored the same piece differently from one day to the next). What interests me, is what it reveals about how readily people accept its verdict; which is neither temporary, nor new.
It’s the same pattern I keep finding whenever I sit long enough with something that annoys me. It’s the same pattern that resurfaces in almost every piece of “verwondering“, every honest bout of wondering-out-loud, I’ve written.
People don’t dislike AI-assisted writing because they’ve studied its statistical fingerprint; they dislike the idea of being fooled. And a tool that hands them a percentage lets them stop thinking at exactly the point where thinking was required. The number becomes the argument and the structure becomes suspicion. The people most certain they can spot it are, more often than not, applying a rule they’d fail themselves: the same reader who calls a well-organised paragraph “obviously AI” has never noticed that the essay they admired last year, the one that moved them, was built on the same devices. And so, they are not defending human writing, rather, they are defending their own sense of having caught something, which is a much older and much less flattering habit than it looks like from where they stand.
None of this means the concern is invented or invalid. Plenty of what fills a social media feed now is genuinely hollow, generated at a volume no editor could keep up with, and the instinct of wanting some way of telling the difference is a reasonable one. But wanting a filter and trusting the very first filter that’s offered to you are not the same thing. And having such filter in the first place is something nobody seems to be thinking through.
A tool flagging a piece at sixteen percent at a first run and, after my adding of two sentences to a paragraph, getting a re-run percentage rising to double that score, while also flagging another piece at a hundred percent regardless of what was removed, is not measuring authorship. It’s measuring density of pattern, and mistaking that for a verdict on the person who made it.
So, here’s what I concluded, for now, because I don’t think this deserves a clean ending any more than the questions it raises did. I’m not going to strip the em-dash, or flatten a tricolon. Nor leave a rougher sentence in on purpose, to satisfy a reader I’ll never meet who has already decided what proof looks like. Because that’s not writing for a reader anymore; it’s writing for a suspicion. And a suspicion is an even worse audience than a slogan, because at least a slogan admits it isn’t thinking.
What I will do is notice when a piece is built entirely out of architecture, with no loose ground left for a reader, human or otherwise, to stand on, and ask whether that’s serving the idea or just proving that it can be built.
Form follows function. But only when I decide what the function is.
https://www.linkedin.com/pulse/guilty-structure-roland-biemans-fpcee
Comments are closed