Diagnosis · any draft (procedure, not diagnosis)
How to not sound like AI when writing
Every list of things to avoid treats every tell as equally likely. They are not, and fixing them out of order is how an edit ends up worse than the draft.
To stop sounding like AI, edit in the order the patterns actually occur, not alphabetically. Cut the validating opener first, since it shows up in nearly every model's output. Un-bullet anything casual next. Check vocabulary last, because word choice is the least reliable signal, and skip single em dashes and rule-of-three lists entirely. Our own research found both are common in writing no model touched.
Why this happens to any draft (procedure, not diagnosis) in particular
Most advice on this topic reads as a flat list: avoid a stock adjective, avoid the em dash, avoid rule-of-three phrasing. Applied in that order, an edit spends its first pass on low-yield targets and often removes things that were never the problem. Our own social-sentiment review found that several commonly cited AI tells, including the lone em dash, one specific overused verb, curly quotes, generic office jargon, three-item lists, and title-case headings, do not hold up under scrutiny; they are common in writing no model touched, and stripping them just flattens a draft's rhythm for nothing. Meanwhile the patterns that are actually universal across the models we logged sit further down most checklists, or do not appear on them at all. Editing in frequency order fixes the loudest signal first and leaves the disputed ones alone.
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
- Fix this first. It is the closest thing to a universal signal in our logs, and cutting it changes the whole opening of a paragraph in one move.
- bold-header-list
- Fix this second. Removing the bold-label bullets from a casual answer strips the report shape a chat reply defaults to, the next most common structural pattern.
- rule-of-three-list
- Leave this alone. Our own social-sentiment review found three-item lists refuted as a tell, common in writing no model touched. A parallel triplet is rhythm, not proof something generated it.
- single-em-dash
- Leave this alone too. The same review found the lone em dash refuted, common in writing no model touched. A run of two or more in one paragraph is the real signal; a single dash on its own is not.
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 Steady Hand, plain and sequential: states the step and stops.
Of course! Here's a quick guide to staying hydrated — it's not really about willpower, it's about habit — and once the habit sticks, the results follow fast and stay that way. Drink a glass first thing in the morning — before coffee, before anything else — and keep a bottle within reach the whole day. Let me know if you'd like a full daily schedule broken out by hour.
Plain: put the water by your bed tonight. Drink the whole glass before you fix breakfast or pour the coffee. Fill a bottle in the morning and keep it where you can see it — desk, kitchen counter, bag handle. Drink from it across the day. That's the habit. Once it's automatic, you don't think about it, and you stay ahead of thirsty. The schedule by hour might help if the first step doesn't stick, but you can test this one first.
How to fix it yourself
- Cut the validating opener Delete any sentence or fragment that agrees with, praises, or grades an implied question before it adds information. Start the paragraph on the first sentence that actually says something.
- Turn casual bullets back into sentences If a list answers something you would say out loud in one breath, write it as a sentence instead. Keep real bullets only for something you would actually hand someone as a checklist.
- Check vocabulary last, sparingly Scan for the handful of words that inflate significance without adding a fact, the kind that dress an ordinary claim up as a verdict. Replace each with the plain version of the claim, and stop there.
- Read the edited draft against the original If the edit reads flatter or less like you than the draft did, you likely cut something that was a personal habit rather than a shared AI pattern. Put it back and move on.
Questions
How to avoid AI while writing?
Write your own draft first if you can. If you are editing an AI draft instead, do not scan alphabetically for banned words. Fix the structural habits first, meaning validating openers and bullet scaffolding, since those are the patterns nearly every model shares, and leave the disputed vocabulary rules alone.
How to avoid sounding like chatgpt?
Chatgpt-specific habits worth checking first: an opener that praises the question before answering it, a closing offer to keep revising, and bullets applied to casual topics. Cut those three and most of the chatgpt-voice complaint disappears, without needing to touch vocabulary at all.
What makes writing sound like AI?
A small set of structural habits, not a banned-word list: an opener that validates before it informs, headers and bullets on casual topics, and constant direct address. All three showed up across nearly every model in our own logs, which is why fixing them first does most of the work.
How to make something sound like it's not AI?
Fix the structural tells in frequency order rather than chasing a word list, and stop once the flat, coach-like tone is gone. Over-correcting past that point, stripping every em dash and every list of three, tends to make an edit read worse than the draft did.
How to ask chatgpt to not sound like AI?
You can ask, and it will change some surface habits for that one reply, but the structural defaults return in the next one. Editing the output after the fact, in frequency order, holds up better than a standing instruction the model drifts away from over a longer conversation.
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.