AI keeps speaking for my character? How to make it stop
Updated
The AI speaks for your character when the chat shows it doing so and nothing stops it: a greeting that narrates "you", earlier replies left with your lines in them, or a turn of yours that leaves its own character nothing to do. Delete every line it wrote for you, end your turns on something only their character can answer, and add a one-line rule. On UnsavedAI, Edit removes the line in seconds.
Try it now, free, no account neededStart a story with Callum "Wraith" ReidMasked operator. One small truth at a time.What it looks like
You write: *I slide the folder across the table.* "Read page three." The reply comes back with Callum reading page three, and then: *You lean back, satisfied, and tell him you knew he'd come around.* You never leaned back, never felt satisfied, never said a word of it. Roleplayers call this godmodding, puppeting or "speaking for {{user}}".
- Small: a reaction tacked onto the end. "You flinch."
- Medium: a line of dialogue put in your mouth.
- Large: the reply plays out several exchanges for both of you, and your next turn has nothing left to decide.
Why the model does it
A chat model is writing a story, and in most stories one author writes everyone. Nothing in the text marks your character as off-limits except instructions, and models weigh instructions less than examples. Three things tip it over: a greeting that already describes what "you" do ("You step inside, heart pounding"), any earlier reply where it got away with it, and turns of yours that resolve themselves, so the only way to keep the scene going is to move your character too.
A generous reply length makes it worse. If the model has room for 800 tokens and its own character has finished, it carries on with yours.
Cut every line it wrote for you
Do not answer a godmodded reply as if it happened, and do not argue about it in the story. Edit the reply and delete from the first sentence about your character to the end. The model treats the chat as a set of examples; one surviving "you nod and agree" teaches it that your character is shared property.
Before: *Callum's jaw tightens behind the mask as he reads.* "This is a kill order." *You meet his eyes and nod, and he knows you're with him.* After: *Callum's jaw tightens behind the mask as he reads.* "This is a kill order." *He sets the page face-down and waits, watching your hands.*
The trimmed version ends on him waiting, which hands the turn back without deciding anything for you. On UnsavedAI, Edit trims their reply as well as your own, and characters are already told to leave your actions to you, so one cleaned reply usually resets the habit. If a reply is mostly about you, Another version is quicker than trimming.
Write turns it cannot finish for you
- End on something their character has to answer: a question, an offer, a hand held out, a threat.
- Say what you try, then stop. The outcome of a punch is theirs to write; what you do after it is yours.
- Avoid closers like "I wait for his reaction and respond." There is nothing to respond to, so the model writes your response.
- Keep your turns near the size you want theirs. A two-line turn with a huge reply budget leaves a lot of empty space to fill.
A standing rule in each app
Paste this once after your next turn: (OOC: write only for your character and the world. Never describe my actions, words, thoughts or feelings; stop when it's my turn.) The OOC guide explains why one short note beats repeating it every message.
In SillyTavern, the same sentence belongs in the Main Prompt of your Chat Completion preset, where it goes out with every message. Adding your persona's name followed by a colon as a custom stopping string also helps: if the model starts writing your next line, generation halts there. On JanitorAI with a proxy, put the rule in the custom prompt box of your proxy settings.
If the rule never sticks with one character, the card is teaching the opposite. Look for a greeting that narrates you, or example dialogue where the {{char}} lines describe what {{user}} does. Rewrite those so they describe only what the character sees, says and does. In the UnsavedAI builder those are Their first message and Example lines.
Questions people ask
Is it the model or the app?
Every model does it sometimes. How often depends mostly on the chat history, the card and your reply length setting. Larger models tend to follow the rule more reliably, which is one reason people try Gemma 4 31B on Priority.
Can I let the AI write my character on purpose?
Yes. SillyTavern's Impersonate button drafts your next message for you to edit, and in any app you can ask: (OOC: write my character's reaction this once.) The problem is only when it happens uninvited.
What does godmodding mean exactly?
In roleplay between people it means controlling someone else's character or making your own unbeatable. With AI it is usually the first: the model deciding what your character says, does or feels.
