Pangram verdict · v3.3
We believe that this text is a mix of AI and human-written content.
AI likelihood · overall
MixedArticle text · 808 words · 6 segments analyzed
I finished writing an article recently. I did what I always do before I queue something to be published: read it aloud, as if reading a storybook to a child. It's how I find the rhythm: where a comma should sit, whether an em dash earns its beat, and which parenthesis a reader can skip without missing much.Lately, though, the read-through has turned into a hunt. I take out em dashes and replace them with commas or full stops, keeping the rhythm steady while making the change less obvious. I look for words like "genuinely", "honest", "quiet", "silent", or "worth", and swap them for something more neutral, almost as if I'm burying my own style. Those are just some of the words saved in a file on my desktop, called the Literary Graveyard. It's full of familiar words, phrases, and patterns that make people read a piece as AI-generated rather than crafted. I found them in arguments in Slack channels, LinkedIn posts, and firsthand comments from a former boss, who was adamant that I strip the em dashes out of what I was writing, grammatically correct as they were, just in case someone assumed it was written by AI.What irks me most is that they were mine first. I, like many others, wrote like this for years before any model did. A machine learned those words, structures, and patterns from people like me, and it repeated them back to me until they stopped feeling like my own. That's a lot of self-surveillance for someone who just wanted to write about design tokens.It's a loop I recognise from years spent in product and design systems.
I've watched it play out plenty of times. You put a pattern into a system, people copy it, and after a while the copying is the proof: it's everywhere, so it must’ve been right. Sometimes the system is cataloguing something that already worked. Other times it invents a behaviour out of nothing, then treats its own spread as the verdict. Half the job is tracing a pattern back to work out which of the two it was, whether anyone actually chose it or it just piled up. I see the same loop in language now, except there's nothing to audit it against.
I can't tell which direction the influence runs anymore.The words came from us firstWhen I first learned about AI, I assumed models invented their own vocabulary. They don't, not really. Researchers at Florida State wanted to understand why ChatGPT uses the word "delve" so frequently. The culprit turned out to be the feedback stage, where humans rate the model's answers to teach it what "good" looks like.When researchers tested it, readers rated the word "delve" lower than any other buzzword.
So this wasn't the model copying a word people loved. A specific group of human reviewers, at one moment in training, nudged it toward "delve", and the model learned the lesson. The words we now roll our eyes at started as a snapshot of what one room of people rewarded.
They came from us first.Then it came back.Researchers at the Max Planck Institute listened to more than 737,000 hours of unscripted podcasts and were able to tie the rise of words like "delve" directly to the release of ChatGPT. If you spend enough time talking to a model, its words begin turning up in your own vocabulary.The influence had reversed. We taught the model, and then the model was teaching us.And then it turned back once more.Once "delve" became a punchline, people used it less frequently than they had before ChatGPT existed. In fact, it was used so infrequently that newer models were tuned to stop using it at all. We started editing ourselves so we wouldn't sound like a machine, while the machine was edited so it wouldn't sound like itself.My Graveyard file sits right in the middle of this. Every entry in it is me editing my own writing because of something a model does. The same move those podcast hosts made with "delve", only pointed the other way.Toward the middleThere's a name for where this ends up. Model collapse. In 2024, Nature wrote a paper on what happens when models train on the output of other models.
The rare, specific details are the first to disappear. Round after round the unusual gets averaged away, and by the ninth round the thread is gone entirely.If you work in design systems, you'll know this by another name. We call it drift — when a system slowly stops meaning anything, because every team keeps bending it into their own one-off versions until there's no shared standard left. What I'm describing runs the other way — instead of scattering, everything pulls toward the middle. The rare and the particular get sanded off, and what's left is the most ordinary version of that thing.