Lunelle ← Back to home
Guides

How to stop an AI character from speaking for you

Updated

An AI character speaks for you because the model behind it is continuing a story, and in the novels and stories it learned from, one writer controls every character. It is more likely to write your lines when your turn leaves a gap, when earlier messages already include your character’s words or actions, and when the only guard against it is a rule that says what not to do. To stop it, cut those lines out as soon as they appear, end each of your turns on something the character has to answer, keep your character out of the opening and the example dialogue, and word any rule as what the AI should do.

These fixes work in Character.AI, Janitor AI, SillyTavern, AI Dungeon, ChatGPT, and other apps. A few apps also have a setting that ends a reply as soon as it starts writing for you, covered near the end.

What counts as speaking for you

Speaking for you means the AI makes a choice that belongs to your character: what they say, what they decide to do, or what they feel. Roleplay communities have names for this, such as godmodding and powerplaying, and it is a long-standing complaint in shared storytelling. With an AI partner it usually shows up at the end of a reply, after the character has done their part, as a line or two about you.

Not every sentence with “you” in it crosses the line. A narrator has to describe what happens to your character, especially in text adventures, where the whole story is told in the second person. These examples run from the narrator’s normal job to a clear case of the AI taking over:

Justin Alexander, a game designer who writes about running tabletop games, gives a rule that fits all of these: never assume a character does something that needs a meaningful choice unless the player makes that choice. Most players want the AI to stay out of their words, decisions, and feelings, and accept that the narrator describes what happens to them. Small actions are a matter of taste. Decide where your line is, because you’ll write it into a rule later.

How human roleplayers handle it

People who wrote stories together on forums and in chat rooms had the same problem with each other long before AI, and their communities wrote rules for it. The rules come from one idea: each player controls only their own character, so any action that involves someone else’s character stops halfway and waits for that player.

RP Repository, a long-running roleplay community, explains it with a hug. If your character goes to hug someone, you write the reach and end your post there, because the other player might have their character pull away. Forum Roleplay’s guide to bad roleplay applies the same rule to fights: you write the swing, and your opponent writes whether it lands. Words like “tries” and “aims” keep the result open for them.

Tabletop games split the work the same way. The introduction to the D&D rules describes the Dungeon Master setting the scene, the players deciding what their adventurers do, and the DM determining and narrating the results.

Both ideas carry over to AI roleplay. You can ask the AI for the same division of work, where it plays the world and every character except yours. You can also follow the forum players’ rule in your own turns, ending on an attempt or a question so the other side has something to answer.

Why the AI writes your lines

An AI model writes a reply by predicting, word by word, the text most likely to come next. It has no fixed sense of whose turn it is beyond what the conversation shows. Several things in a conversation push that prediction toward your character, and a chat where it keeps happening often has more than one at once.

It’s continuing a story

To the model, a chat is one long piece of text. Hugging Face’s documentation for its Transformers library describes chat models as language models that continue a sequence of text, with hidden markers added to show where each message starts and ends. Without the marker that opens the AI’s turn, the documentation notes, a model may carry on writing the user’s message instead of replying to it.

The markers end the AI’s turn, but inside that turn the model is writing prose. Murray Shanahan and colleagues, in a 2023 paper on role play with language models published in Nature, point out that the text these models learn from includes novels, screenplays, and biographies. In a novel, one narrator writes every character in the room. Prose that narrates one character’s actions can easily go on to narrate the next person in the scene, and in a roleplay that is often your character.

Your turn left a gap

The model has to move the story forward with every reply. If your message gives the character nothing to respond to, the next thing that can happen is often something your character does. Janitor AI’s help center describes this: when you write only dialogue, without your character’s actions or reactions, the AI may step in to complete the story and fill the space you left.

I walk into the tavern.

A reply to that turn might read:

The tavern is warm and loud, and the smell of roast meat hangs over the tables. You make your way to the bar and order an ale, scanning the room for anyone who matches the description.

Nothing in your message gave the scene a reason to wait for you, so the AI decided what you do once you’re inside. A turn that ends on something only the other side can answer, such as a question to the barkeep, leaves the AI a job of its own.

Your lines are already in the chat

Models continue patterns. Shanahan and colleagues describe this as the ability to carry on in the same vein as the text so far, and it’s how a character keeps its voice from one reply to the next. It also means mistakes get copied: if earlier replies included your character’s words, the model treats that as how this story is written, and each reply that keeps them adds another example.

SillyTavern’s documentation makes the same point about chat history: the model picks up word choice and writing style from earlier messages, and once a chat has some history, changing the instructions has a limited effect. The first examples in any chat are the character’s greeting and, if it has them, its example dialogues. Character.AI’s help center calls dialogue examples one of the most effective tools in a character’s definition. A greeting that says “You smile and take her hand” teaches the model, from the first message, that the narrator writes your actions.

The rule says what not to do

The usual fix is a line in the character’s definition such as Never speak for {{user}}, and it often doesn’t work. Janitor AI’s own help center says these instructions “just don’t work the way people think they do,” and suggests that a rule about your character can lead the AI to write about your character more. SillyTavern’s documentation says the AI follows instructions about what it should do more easily than instructions about what it shouldn’t.

The companies that make the models give the same advice for prompts in general. Anthropic’s prompting guide for Claude says to tell the model what to do instead of what not to do, and OpenAI’s prompt engineering best practices say the same. A 2022 study by Joel Jang, Seonghyeon Ye, and Minjoon Seo adds some evidence, with a caveat: language models did worse on tasks phrased with a negation, and larger models did worse than smaller ones. Those were older models on simple test tasks, so the study doesn’t show that today’s chat models ignore every “don’t”. It does fit the practical advice to state the behavior you want.

Some models do it more than others

Models differ in how readily they write for the player, depending on how they were trained. AI Dungeon’s help pages show this within one app. They describe one model, Wayfarer Small, as popular for keeping the AI from deciding what the player does, and give other models, Muse and Harbinger, a specific line to add when they start writing for your character. If your app lets you choose the model, switching is a fair fix to try.

What to do when it happens

Deal with it in the reply where it first appears. Every reply that stays in the chat becomes part of the example the AI follows, so a line about your character that you leave in makes the next one more likely.

  1. Cut the reply at the point where it starts writing for you. Most apps let you edit the AI’s messages, so keep the character’s part and delete the rest.
  2. If the whole reply is built around a choice you didn’t make, regenerate it, or rewind the chat to before it. Character.AI calls this Rewind to Here.
  3. Don’t play along in the story. If you carry on as if the AI’s decision happened, it stays in the history the AI copies.
  4. If it keeps happening, add one short out-of-character note, then carry on in character. See how to steer an AI story for the format.

Here is a reply that needs a cut:

Mara crouches beside the body and turns the dead man’s hand toward the light. “Ferry ticket. Dated tomorrow.” She glances up at you. “Thoughts?” You shrug and tell her it looks like a setup, then head for the door to find the ticket office.

Everything after “Thoughts?” is yours to write. Delete it, and the reply ends on Mara’s question. SillyTavern’s documentation advises never letting the AI get away with something you don’t want and never carrying on as if a bad reply were fine. An edit takes a few seconds, and it removes an example the AI would otherwise copy.

How to end your turn

Your own messages make up half of what the AI reads, so they are the best place to prevent the problem. Two habits do most of the work: write your character’s part in full, and stop where the other character has to respond.

Writing your part in full means saying what your character does, says, and wants, and how they react. Janitor AI’s help center recommends this for the reason above: the fewer gaps you leave, the less the AI has to fill. A short paragraph is enough.

Stopping where the other character has to respond is the forum players’ rule from earlier: write the attempt, and leave the result and the reaction to the other side. Good endings are a question to a character, an action that needs a reaction, or something your character shows or offers.

I look at the ticket.

This turn leaves the next move open, and the AI may take it by deciding what your character notices, what they conclude, and where they go next.

I turn the ticket over. There’s a harbor-office stamp on the back, dated this morning. “Somebody bought this today, Mara. Who sells a ticket to a man who’s already dead?”

This version says what your character does and thinks, and ends on a question only Mara can answer.

The same applies to the other character. When your turn decides how they react (“I tell a joke and she laughs”), the chat starts to show both sides being written in one turn, and the AI may follow that pattern. Leave the other character’s reactions to the AI.

Fix the greeting and the example dialogue

If you made the character, check its greeting and example dialogues first, since they are the earliest examples in every chat. They should contain the character’s actions, the character’s words, and reactions to you, but no actions by you. Glub’s character guide, a bot-making guide for Janitor AI, gives this rule: write the character’s reactions to the user, not actions by the user, unless you want writing for the user to become the bot’s habit.

A greeting that writes your actions:

Mara looks up as you walk in. You shake the rain off your coat and kneel beside her, studying the body while she shines her flashlight on the dead man’s hand.

The same scene with the action left to Mara:

Mara looks up as the door opens. “You’re late.” She tilts her flashlight toward the dead man’s hand: a ferry ticket, dated tomorrow. “Tell me why a man buys a ticket for a trip he knows he won’t take.”

In the first version your character walks in, shakes off the rain, and kneels, so from the first message the narrator is writing your actions. The second gives Mara the action and the dialogue, and ends on a question addressed to you.

In example dialogues, keep each side to its own lines. The user’s turns contain only what the user says, and the character’s turns contain only what the character says and does. Character.AI’s help center also suggests giving the user a different stand-in name in each example exchange and marking where each exchange ends, so the model reads them as separate examples instead of one long conversation. Our guide to writing an opening scene covers greetings in more detail.

If you’re playing someone else’s character and its greeting writes your actions, edit the greeting in your own chat if your app allows it. If it doesn’t, write your first turn as a full action of your own, and edit the first reply that copies the greeting’s habit.

How to word a rule that helps

Add a rule once the habits above are in place, since a rule on its own is the weakest fix. It helps most when it says what the AI should do, names your character, says where your character’s part begins, and sits where the app sends it last.

Placement matters because models give extra weight to text near the end of what they read, as Chub’s documentation notes. Many apps let you put instructions just before the model writes its reply: SillyTavern and Chub call these Post-History Instructions, SillyTavern also has an Author’s Note, and AI Dungeon sends its Author’s Note toward the end of every input. SillyTavern’s documentation says the AI usually gives instructions placed after the chat history a higher priority than instructions at the start. A line at the top of the character’s definition has to compete with everything written since.

A rule you can adapt:

Write only for Mara and the other characters in the scene. Riley is my character, and I write everything Riley says, decides, and feels. You may describe what happens around Riley and how the world reacts to what Riley does. End your reply when Riley needs to speak, choose, or act.

The first sentence gives the AI its role in positive terms. The second names your character and draws the line from the first section of this guide; add “and does” if you want small actions left to you as well. The third keeps the narrator’s normal job, so the AI doesn’t stall the scene, and the last tells it where to stop. Replace the names with your own. In apps that use the placeholders {{char}} and {{user}}, such as SillyTavern and Janitor AI, you can use those instead. SillyTavern’s documentation also mentions a shorter version that some users prefer: Write {{char}}'s responses in a way that respects {{user}}'s autonomy.

AI Dungeon’s own suggestion for its Muse and Harbinger models is a short line in the Author’s Note, such as ”- don’t write dialogue or actions for NAME”, with your character’s name in place of NAME. It is phrased as a negative, but it is short, names the character, and sits at the end of what the AI reads. If a positive rule doesn’t work with your model, try the wording your app recommends.

Keep it to one short rule. Character.AI’s help center warns that a long list of rules gives the AI more chances to get confused and apply the wrong one.

App settings that stop it

Some apps can end a reply the moment the AI starts your turn. They use a stop sequence: a piece of text that ends the reply as soon as the model writes it. Set to your character’s name, it stops the reply when the model starts a new line of your dialogue, before you see it.

A stop sequence only catches your character’s name used as a label, as in a chat log. It doesn’t catch narration such as “Riley nods” in the middle of a paragraph, so it works best alongside the habits above. If the AI writes about you mostly at the end of long replies, a shorter reply length can also help where your app has that setting, because it leaves less room after the character has done their part.

When nothing works

If the AI still writes for you after edits, better turn endings, and a rule, the chat history is probably the cause: too many earlier replies show the AI writing your character. Start a new chat and carry the story over with a short summary of what has happened so far. Janitor AI’s help center gives the same advice for chats that have gone off course. See why AI characters forget for what a good summary keeps.

If it happens in every new chat with the same character, look at the greeting and example dialogue again, or try another model if your app lets you choose one.

Common questions

Why does Character.AI keep talking for me?

For the same reasons as in any other app: a gap at the end of your turn, a greeting or example dialogue that includes your actions, or earlier replies that wrote for you and stayed in the chat. Edit the reply, swipe for a different one, or use Rewind to Here to go back to before it started. If you made the character, check its greeting and definition for lines that act for the user.

Does a “don’t speak for me” rule work?

Sometimes, but it is the weakest fix on its own. Janitor AI’s help center says these instructions often fail, and SillyTavern’s documentation says rules that state what to do work better. No rule outweighs a chat history full of the AI writing your lines, so fix the history and your turn endings first, then add a positive rule where your app sends instructions last.

Is the AI godmodding?

In roleplay terms, yes. Godmodding and powerplaying mean controlling a character that belongs to someone else. The AI isn’t doing it to win, though: it is continuing the story the way the text so far suggests. See our glossary of roleplay terms.

Can I let the AI write my character sometimes?

Yes. Some players are happy to let the AI handle small actions or move their character through a quiet scene. Say so in your rule, for example: “You may describe Riley’s small movements, but leave Riley’s words and decisions to me.” SillyTavern and Chub also have an Impersonate button that drafts your next turn for you to edit.

In Lunelle AI you play adventures and ongoing stories in your own words, and characters keep their personality, voice, and established facts over long plays. The habits in this guide work there as they do in any app. See what Lunelle AI is.

Sources