Diagnosis · any draft (general head page)
Why does my writing sound like AI?
You did not run it through a checker. You noticed it yourself, mid-sentence, and that instinct is usually right.
Writing reads as AI-generated because of a small set of patterns that show up whether or not AI touched it: a validating opener before the point, headers and bullets imposed on a casual thought, and a flat, coach-like tone that talks at the reader instead of with them. Across twelve models we logged, all three appeared in every single one, tied at the top rather than ranked. Check for those first.
Why this happens to any draft (general head page) in particular
Most AI-sounding prose was not built as an essay or a post. It came out of a chat window, one reply in a back-and-forth, and it kept the shape of that context after being pasted somewhere else. A chat reply defaults to opening with agreement, because the assistant is responding to a prompt and validating it is the safest first move. It defaults to bullets and bold labels, because a list is legible mid-conversation even when the topic was personal. And it defaults to addressing you directly and often, because the assistant's whole job in that window is to talk at one reader. Move that reply into an email, a post, or a page, and the conversational scaffolding stays, uninvited.
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 opener that praises or agrees with an implied question before saying anything showed up in all twelve models we logged, almost always as the reply's first clause, a habit that survives being pasted anywhere.
- bold-header-list
- Bold-label bullets appeared in all twelve models too, even on casual one-on-one topics; the assistant reflexively imposes report structure on a plain paragraph.
- testament-family
- When a model reaches for significance instead of a fact, this is the shape it takes: a copula dressed up as a verdict.
- vague-attribution
- A claim needs a source, and a chat reply with none handy borrows the weight of an unnamed study instead of pointing to an actual one.
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 Essayist, calm and precise: the writer nearly invisible behind the subject.
Great question! Remote work is not just a trend, it's the future of how we work. Studies show that employees are more productive when given flexibility. Here's what you need to know: - **Flexibility:** You can work from anywhere. - **Productivity:** You'll get more done without a commute. - **Balance:** You'll finally have time for what matters. Let's dive into how you can make the switch today!
Remote work is not a passing experiment nor a temporary accommodation. It is how more people now work, and evidence suggests that employees given flexibility do not slow down; they perform better. Freedom from a commute, the ability to choose one's own setting, the room for a life that is not split neatly between desk and home — these are not perks but the structure itself. How one makes the switch is a matter of temperament and circumstance as much as logistics.
How to fix it yourself
- Read the first clause of every paragraph Cut anything that grades or agrees with an unstated question before it says something new. If the first six words could be deleted without losing information, delete them.
- Un-bullet the casual parts Any list that answers something you would say out loud in one breath should go back to a sentence. Keep bullets only where you would actually hand someone a checklist.
- Trace every claim to its source Replace a sourceless appeal to research or experts with the actual study, your own experience, or delete the claim. An appeal with no source behind it is the fastest way a reader stops trusting the rest.
- Read it out loud once The coach-like second-person address a chat reply defaults to sounds fine silently and strange out loud. If you would not say a line to the person's face, cut it.
Questions
Why does my writing sound like AI when it's not?
Some of the patterns models overuse, like a validating opener, bullet-and-bold scaffolding, and constant direct address, are also just common habits from reading a lot of chat-style content, work Slack, or LinkedIn. You can pick them up without touching a model. The fix is the same either way: cut the habit, not the accusation.
What makes writing sound like AI?
In our own logs across twelve models, three patterns were nearly universal: an opener that validates before it informs, headers and bullets applied to casual topics, and a flat, second-person-heavy tone that addresses the reader constantly. None of these need vocabulary swaps to fix; they're structural habits, not word choices.
What words sound like AI?
Fewer than people think. Our own logs found the classic essay-mill vocabulary largely absent from casual assistant replies. What's actually universal is structural: the opener, the bullets, the second-person density. Chasing a banned-word list misses where the real signal is.
Why does all my writing look like AI?
If every piece gets flagged regardless of topic, check for a habit rather than a phrase: do you open by agreeing with an implied question, reach for bullets on anything with more than one part, or address the reader directly in most sentences? Those three habits generalize across everything you write.
Can I check if my writing sounds like AI?
By eye, once you know the specific patterns: the validating opener, the bullet reflex, the flat second-person tone. Reading for those three catches most of what makes a draft sound machine-written. If you want a second pass, underslop marks any of those patterns it finds when you paste a draft in.
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.