Diagnosis · other people's writing / the open web
Why everything sounds like ChatGPT now
You are not imagining it. Reviews, comments, captions, and cover letters are converging on one register, and it is not because everyone is quietly pasting from a chatbot.
Everything sounds like ChatGPT because the convergence is behavioural, not lexical: a handful of discourse moves, a validating opener, near-zero hedging on uncertain claims, a cheerful close, that our own twelve-model logs found in most systems tested. Those moves read as competent and safe, so writers copy the register the way office email once copied a stock phrase, without anyone touching a chatbot.
Why this happens to other people's writing / the open web in particular
The open web is not one genre, which is exactly why the convergence shows up there and nowhere narrower. A reader now meets model output dozens of times a day without asking for it: an autocomplete finished a sentence, a support reply was drafted by an assistant, a friend pasted a chatbot's phrasing without saying so. Each contact nudges the reader's own ear a little further toward the register, the way an accent spreads through contact rather than through a lesson. It also helps that the convergence spans more than one company's models: our twelve logged conversations cover eight labs and roughly ten distinct base models (three of the twelve are the same DeepSeek engine run under different inference settings, not three separate labs), and the same handful of discourse moves turned up across nearly all of them. A shared training target produces a shared output shape from different starting points, and that shape then leaks into writing no model ever touched.
The tells that show up here
Not every AI tell appears everywhere. These are the ones that cluster on this surface, and what it is about the format that produces them.
- chatbot-artifact
- The open web absorbed the validation opener first because it is the move readers meet most often: a phrase lifted from a support chat into a comment reply where nobody asked a question.
- not-x-its-y
- The negative-parallelism reframe showed up in eight of our twelve model logs and reads as a punchy, quotable move, so human copywriters and captions kept the shape long after the underlying phrase stopped sounding machine-made.
- throat-clearing
- A pivot phrase signals the real point is about to arrive, which is exactly the job it does in a model reply after a long hedge. Human writers reach for it as a shortcut to sound like they are finally getting real.
- em-dash-chain
- Nine of twelve models append a clarifying or pivoting clause mid-sentence with an em dash instead of stopping. Screenshotted chatbot replies made the habit visible, and it migrated into captions and posts that never touched a model.
Before and after
A generic AI draft on the left. On the right, the real output of running that draft through underslop, checked so it adds no name or number the draft did not already carry. The voice is The Porch Regular, plainspoken and unhurried
Great question! Honestly, remote work isn't just a trend, it's a fundamental shift in how we think about productivity. Studies show that flexible schedules boost morale and output. At the end of the day, it's not about where you work, it's about how you work. If you're on the fence, here's the thing: give it a real chance before you judge it. You've got this!
Well, I wouldn't call remote work a trend exactly. More like people finally noticing the fence was already down. You can work from the kitchen table or from an office downtown, and either way the work gets done or it doesn't. Studies say flexible schedules help with morale and output, and I figure that's probably true. It ain't really about where you are. It's about how you do what you do. So if you're thinking about trying it, you might as well give it a fair shake before you decide. Can't hurt to see.
How to fix it yourself
- Read your last five sent messages back to back Look for the same opening move each time: a validating first line, an exhaustive list, a cheerful sign-off. One instance is a habit. Five in a row across different topics is the register talking, not you.
- Check your second-person density Count how many sentences address 'you' against how many use 'I' or an inclusive 'we.' Our logs show model output leans heavily toward 'you' with almost no 'we,' a coach's stance rather than a conversation.
- Look for the missing hedge On a genuinely uncertain topic, a big decision, a forecast, check whether you stated it flatly. Confident, hedge-free language on real uncertainty showed up in ten of our twelve model logs; real uncertainty usually sounds uncertain.
- Send the paragraph through underslop before you post it Paste up to 500 words into underslop and set a voice slider. It marks the specific moves above instead of guessing, and shows what is left once the borrowed register comes out.
Questions
Why are people starting to sound like ChatGPT?
They meet the register constantly, in autocomplete, in support replies, in pasted chatbot answers, and pick it up the way any dialect spreads through contact. It also reads as competent and safe, so writers copy it on purpose, the same way office email absorbed a stock phrase long before anyone opened a chatbot.
Why does ChatGPT sound like a real person?
It doesn't, exactly. It sounds like a specific kind of person: warm, hedge-free on hard questions, addressing you directly, closing with encouragement. That combination is rare in real writing but common across the twelve model logs we reviewed, which is what makes it recognizable rather than generically human.
What makes it sound like AI?
A small set of discourse moves repeated across nearly every model we logged: a validating opener, structural scaffolding even for casual topics, suppressed hedging on uncertain claims, and a closing question or cheer. No single word gives it away. The pattern is the tell, not the vocabulary.
Why does AI sound like that?
Because the training target rewards a helpful, thorough, unfailingly calm answer, and our twelve logged conversations span eight labs and roughly ten distinct base models that converged on the same shape to hit it. The register is an artifact of the reward signal, not a style anyone chose, and it shows up the same way across models built independently of each other.
This page explains the cause. underslop is the edit: paste up to 500 words on the free tier, and it marks the tells it could not remove. Your text is held for 0 seconds.