AI tell · grade C
ChatGPT write in my voice: why one adjective doesn't transfer
Write like me, casual and a bit witty, is the most common way people ask a model for their own voice, and it usually returns a paragraph with a few contractions and a joke bolted onto the same underlying shape. An adjective describes a texture; it says nothing about sentence length, how much a person hedges, how directly they address someone, or how a paragraph is allowed to end. underslop asks for five calibrated dimensions instead of one word: how outgoing or reserved, how calm or reactive, how organized or spontaneous, how warm or blunt, how inventive or plain. Each one is a slider from zero to one hundred, not a fixed label, mapped from a five-factor personality model rather than picked for flavor. Run the same source paragraph through two different settings on those five sliders and the difference shows up in sentence length and directness, not just in word choice.
What it is
Five dimensions, each a slider rather than a word. Outgoing versus reserved controls how socially present the phrasing is. Calm versus reactive controls emotional temperature. Organized versus spontaneous controls how tightly the flow is structured. Warm versus blunt controls how accommodating the stance is toward the reader. Inventive versus plain controls how much imagery and rhythm the sentence carries without adding anything the draft did not already say. A value within about eight points of the midpoint on any one of the five is treated as untouched, so a slider only shifts the register when it is deliberately pushed toward one end, not by default. This is a mechanism specific to how underslop edits an existing draft, not a claim about how any chatbot follows a style instruction.
How often it shows up
We have not run a controlled test measuring how far a plain adjective instruction shifts a model's output compared to a five-dimension edit; that comparison does not exist in our data and we are not claiming it does. What we can say from our own logs: a coaching-dense second-person address rate held steady across all twelve models we sampled, across every topic in the corpus, and a suppressed hedging rate held across ten of twelve models the same way. Neither number comes from testing persona or tone instructions against the models; both are observations across naturally varying conversations, and they say nothing directly about how far a single adjective would move either rate. What the five sliders change is our own architecture, verifiable in the tool itself, not a measurement of ChatGPT's prompt-following.
Why readers flag it
An adjective describes surface texture. It says nothing about how long a sentence should run, how often a hedge should appear, how directly the writer addresses the reader, or how a paragraph is allowed to close. A model can honor casual by adding a contraction and a joke while leaving all four of those levers exactly where its defaults put them, which is why a nominally casual reply can still read as generated: the word changed, the structure underneath did not.
The fix
Specify more than one axis. At minimum say how outgoing or reserved the address should be, how calm or reactive under pressure, how tightly organized the structure should be, how warm or blunt the stance is, and how much invention versus plain fact the sentence carries. underslop turns those five into calibrated sliders and edits a draft you already wrote toward wherever they land, rather than generating fresh text from a personality description, which is a different mechanism from asking a chatbot to imitate a vibe from scratch.
Examples
This is not just a networking event, it's a chance to meet people who get it. Here's the thing: everyone starts somewhere, and community is what makes the difference. You've got this!
This is a chance to meet people who already understand what you're going through. Everyone starts somewhere, and a good community is what actually helps you keep going.
This is not just a networking event, it's a chance to meet people who get it. Here's the thing: everyone starts somewhere, and community is what makes the difference. You've got this!
This is a chance to meet people who understand what you're working on, not just another networking mixer. Everyone starts somewhere, and having a community around you is what actually moves things forward.
Great question! To sound casual: use contractions, don't be too formal, and add some humor. Here's the thing: doing that alone won't fix the sentence rhythm underneath.
To sound casual, use contractions and drop some formality. A little humor helps too. But that alone won't fix the sentence rhythm underneath.
Questions
Can you train ChatGPT to write in your voice?
Not in a way that persists across sessions from a single conversation. What you can do inside one exchange is give it calibrated instructions about register rather than one adjective, and expect to repeat them, since nothing carries over automatically once the conversation ends.
How do I get ChatGPT to write in my voice?
Name the specific dimensions instead of one adjective: how direct, how organized, how warm, how much invention, how reactive under pressure. A single word like casual underspecifies all five and usually only shifts word choice.
How do I write in my own voice?
The honest test is whether the sentence lengths vary, whether you would say it out loud, and whether the specifics are yours. None of that comes from a personality label; it comes from cutting what a model added by default, the praise-then-answer opener, the cheerleading close, the forced structure.
How do I get AI to sound like me?
Start from something you already wrote rather than a description of yourself. underslop edits a draft you paste toward five calibrated sliders instead of generating fresh text from an adjective list, which keeps your actual specifics in place while the register shifts.
How do I use AI without losing my voice?
Keep the facts and specifics you supplied and treat anything the model added on top as a first draft to prune, not a finished voice to accept. The register comes from what you cut and how you set the dials, not from what the model volunteered.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.