AI Keeps Speaking for You? How to Stop It

Why does the AI keep speaking for you?

The examples in this guide use one scene. A hailstorm hits the glasshouse of a botanical garden at night. You play Sami, a new volunteer, and the AI plays Liesl, the head gardener.

You usually notice it as the AI narrating your character: a line of dialogue you never wrote, or a feeling Sami never had. The model behind a character doesn't think in turns. It reads the chat as one story and writes whatever comes next, and in a story with two people, what comes next is often your character's move. Unless something tells it otherwise, writing Sami's next line seems as reasonable to it as writing Liesl's.

A few things make it much more likely:

• Your message leaves an opening. If it ends with Sami looking at something, stopping halfway through an action, or asking a question anyone could answer, the next beat belongs to Sami, and the model writes it.
• The setup or earlier replies mixed both voices. An opening that described what both characters do, or one earlier reply that narrated Sami and got an in-character answer, shows the model that this is how the story is written.
• Long replies drift. The longer a reply runs, the more likely it goes past the end of Liesl's turn and into yours, usually in the last few sentences.

End each message on something only the AI's character can answer

This is the cheapest fix, and often the only one you need. Before you send, look at your last sentence and ask who would naturally respond to it. If it could be Sami, the model may write him. If only Liesl can answer, because you asked her something or left her a decision that's hers to make, the reply has an obvious place to stop.

Questions work well as long as they're aimed at her and need more than a yes. An action she has to deal with works too, like handing her something or telling her something she didn't know. In the message above, Sami only looks around, so his reaction is the obvious next beat. The rewrite below leaves Liesl a choice only she can make.

Add one line to the setup about who plays whom

If it keeps happening, say it once at the start: which character is yours, which ones belong to the AI, and that its reply ends when your turn comes. One clear sentence in the opening message or character setup is enough. A long list of rules is easier to half follow than a single plain line.

Mention thoughts and feelings as well as words and actions. Those are the parts that slip in most quietly, like Sami's racing heart above. If long replies are part of the problem, add a second line about length.

On tempteo, the roleplay page has three opening scenes you can edit and save as a draft, and nothing is sent until you press send, so you can add this line before the story starts.

Correct it with a short OOC note

When a reply writes for your character anyway, say so right away, out of character, in a line or two. Point at the exact slip, such as Sami climbing the ladder, and ask for the reply again. A specific note works better than a general complaint because the model can see which part to drop.

Keep the tone flat and practical. You don't need to explain or apologize, and a long lecture doesn't work better than one sentence. It just fills the chat with out-of-character text.

Edit or regenerate the reply instead of answering it

It's tempting to reply in character and move on, maybe by having Sami do something different. The trouble is that the bad reply stays in the chat. The model reads its own earlier replies as examples of how this story goes, so every reply that narrates Sami and gets answered makes the next one more likely. One slip is easy to undo. After five, it has become the style of the whole chat.

So take the reply out of the history before you answer. If your app lets you edit replies, cut everything after the AI character's part. If it doesn't, regenerate. Replies are generated and can vary, so a second or third try often stops in the right place with no other change.

If you can do neither, send the OOC note, then write your next message from the last point you accept, as if the Sami part never happened.

Keep your own character's actions in your message

The model fills gaps. If your message leaves Sami halfway through something, it will finish the move for you, and it will pick how he feels about it too. Instead, write Sami's whole move in your own turn, including what he says and where it leaves him, and then hand the next beat to Liesl.

The same goes for asking the AI to narrate your character, even in passing. A line like "describe how Sami reacts" invites the exact habit you're trying to stop, and it tends to carry over into the replies after that.