An AI chatbot speaking for you is probably the most common complaint in roleplay chat, and it comes attached to some of the worst advice on the internet. You write a line. The reply comes back, and somewhere in the second paragraph your character has already answered, already laughed it off, already agreed to the thing you were about to refuse. The scene moved on without you.
Search for a fix and you get the same three sentences everywhere: add a rule. Put "do not write my dialogue" in the description, swipe until it behaves, rate the bad reply one star. None of that is wrong. It's just the weakest lever in the room, which is why people are still complaining about this months after they started doing it.
The stronger levers are in your own messages, and they work on any platform.
What everyone tells you to do
The standard advice has three moves, and they show up in that order almost everywhere.
First, write an instruction into the character description or the system prompt: never speak or act as the user. Second, swipe or regenerate when it happens anyway. Third, leave a rating so the platform supposedly learns.
The first move is the one people actually believe in, so it's the one to examine. A rule feels like it should be binding. You wrote it down, in a field labelled instructions, at the top of the conversation. Surely that outranks everything else.
It doesn't. And the reason has nothing to do with which app you picked.
The model doesn't know there are two of you
Underneath the character card and the avatar, a language model is doing one job. It reads a long block of text and predicts what comes next. Your description, the greeting, every message either of you has sent, all of it arrives as a single document. Nothing in that document marks "this part is a person and this part is the character." There are just turns, and turns look alike.
So when the model reaches the end of your message and starts generating, it isn't weighing whether it has permission to write your line. It's answering a much dumber question: what plausibly follows this? If the document contains fifty turns where a line of dialogue is followed by a reply, and your last turn ended mid-exchange, then the most plausible continuation is the reply. Including yours.
Which is also why the rule underperforms. Your instruction is a couple of sentences near the top of a document that, by message forty, contains thousands of words demonstrating the opposite. Instructions are text. Examples are also text. The examples are winning on volume.
That mechanism sits behind most chatbot behaviour people read as intent, and it's useful well beyond this one failure. There's a plain-English walkthrough of what actually happens between your message and the reply if you want the longer version.
Stop treating it as disobedience and the fixes change shape. They stop being about rules and start being about evidence.
Two failures that look identical
Before changing anything, work out which one you're seeing, because they have different causes.
The first is dialogue theft. The reply contains a line in quotation marks that your character says. This one is almost always about how your turn ended, and it's very fixable.
The second is reaction theft. The bot doesn't put words in your mouth, but it narrates you: your character's face softens, your character steps back, your character feels a flicker of something. No quotation marks, no dialogue, and yet the scene has decided what you did. This one is harder. Plenty of models treat describing the other party's reaction as ordinary scene-writing, and plenty of readers actively like it that way.
Mixing the two together is why generic advice fails. The fix for the first is mechanical. The fix for the second is a preference you have to state and then keep enforcing, since the model has no way to know you object to something most users never mention.
The real lever is how you end your turn
Here's the change that does more than every prompt rule combined. Stop leaving your turn open.
An open turn invites a continuation from your side. A question you asked and didn't answer. A sentence that trails into a decision you haven't made. An action with an obvious next beat, like reaching for a door without opening it. To a text predictor all of those read as an unfinished scene, and the model finishes what's unfinished.
A closed turn ends on something the character has to respond to. You said the thing. You made the decision. You did the action and stopped. The next move is theirs, and there's nowhere for the model to go except into its own character.
Compare two versions of the same beat. "She reaches for the letter, wondering whether to open it in front of him or wait" leaves a decision hanging, and the reply will usually make that decision for you. "She picks up the letter and breaks the seal without looking away from him" is finished. There's no gap to fill.
You can feel the difference within a few messages. It isn't subtle.
Stop demonstrating the thing you don't want
The second lever is what your own messages model. Language models imitate the shape of what they read, and your turns are the clearest example in the conversation of what a turn looks like.
If your messages routinely describe the character's reaction alongside your own action, you've shown that writing the other party's side is normal here. "She sets the cup down and he tenses at the sound" is you writing his half. The model will match you, and now you've taught it the behaviour you're complaining about.
Length asymmetry does the same thing. Write a paragraph-long message with three separate beats in it, and the reply will usually cover three beats too, which means it needs material, which means some of that material is going to be yours. Shorter, single-beat turns leave less room for spillover.
None of this means writing worse. It means noticing that everything you send is a worked example.
Give the character something to do that isn't you
The third lever is scene design, and hardly anyone tries it.
A bot writes your lines partly because you're the only thing in the room. If the whole scene is two people facing each other with nothing else happening, the only content available to a reply is the exchange itself, and the exchange includes both halves.
Put something else in the scene. A task the character is halfway through. Another person walking in. Weather, a phone ringing, a door that won't close, a decision that belongs to them. Now the model has somewhere to spend a paragraph that isn't your character's interior life.
That's the same principle running under most of the advice in the roleplay setup habits that keep a story worth returning to. A scene with texture gives the model somewhere to go. A scene with nothing in it sends it straight back to you.
Where the persona field actually helps
Character descriptions and user-persona fields do have a real effect, just not the advertised one.
"Never write for the user" is a rule, and rules compete badly with demonstrated pattern. A description of who your character is, written as a character, works better, because it stops being a rule and becomes part of the world the model is completing. If the document establishes that your character is guarded and slow to answer, a reply that has her cheerfully agreeing to something now contradicts the story itself, not just a line of instructions nobody is enforcing.
The model isn't obeying you there. It's being consistent, and you've finally given it something to be consistent with.
Most platforms expose some version of this. On some it's a full persona editor. On others, including this AI chat platform, roleplay customisation comes down to choosing a role for the character and writing a context prompt, so that prompt is where the description of you belongs. Fewer fields, and the one field you get carries more weight.
Swiping, editing, and the one move worth making
Swiping is the reflex and it barely helps, because regenerating changes the sample without changing the conditions. Your turn still ends where it ended. The invitation is still sitting there. You're rolling the dice again on the same table.
Editing works, and for the same reason the rule doesn't: it changes the document instead of arguing with it. Delete the part of the reply where the bot spoke as you, keep the rest, continue. Do that two or three times and the recent history stops containing examples of the behaviour.
If your bigger problem is replies circling the same phrases no matter what you write, that's a different failure with a different set of fixes, covered separately in what to do when the replies start looping.
One caution. If you add an out-of-character note asking it to stop, delete the note once it has worked. Leftover stage directions pile up into a pattern of their own.
When you should let it write both sides
Not everyone wants this fixed, and that's fair.
If you're using chat to read a story instead of being in one, a bot that carries both halves is doing you a favour. It moves faster, it fills gaps you didn't want to think about, and you can steer with a sentence every few messages instead of writing a character. Plenty of people use it exactly this way and are perfectly happy.
It becomes a problem only when you wanted agency and got a novel. Not sure which you want? Try one scene each way. Most people find out inside ten minutes.
What this looks like in practice
None of the above depends on a particular app. Friend2Chat's own setup is deliberately small: pick a character, pick the role you want them in, write a context prompt describing the situation and who you are in it. No persona builder, no scene editor, no voice. The conversation is text in a browser, and the interface shows how much the character currently knows about you, so continuity is visible instead of guesswork.
That constraint is exactly why the levers above matter more here, not less. When your only tools are a context prompt and your own turns, how you end a message becomes the main control you have. You can see how the format works by opening any character at friend2chat and starting a scene. The catalogue sits on the homepage at friend2chat.com, sorted by category.
Short version: a bot writing your lines is completing a document that never told it there were two authors. Close your turns, quit modelling the behaviour, put something else in the room, and the problem mostly stops on its own.
FAQ
Why does the AI write my character's lines at all? Because it's predicting what text comes next and has no built-in concept of which turns belong to which person. If your last message left an exchange unfinished, the likeliest continuation is the missing half of that exchange, whoever it belongs to.
Does putting "don't speak for me" in the description work? Sometimes, briefly. It's one instruction competing against every message in the conversation showing the opposite. Typically it holds for a few turns after you add it, then fades as the conversation grows. Changing how your turns end lasts a lot longer.
Is it worse on some platforms than others? It varies with the model and with how the character was written, not with whether you talk to a chatbot online in a browser or through an app. Bots built from example dialogues that blur the two speakers are noticeably worse, which is one of the things to check when you compare platforms, alongside the differences that actually matter between two of the big ones.
Should I swipe, edit, or start over? Edit first, since it's the only one of the three that changes what the model is reading. Swipe if the reply is bad for other reasons too. Start over only when the character has drifted so far that nothing in the recent history is salvageable, which is rarer than people assume.
Do longer messages make it better or worse? Usually worse. A long turn with several beats sets up a long reply with several beats, and the extra material has to come from somewhere. If you want an AI you can talk to as an exchange instead of a monologue, keep your turns to one beat and end them cleanly. That single habit does more than most of the settings people spend an evening adjusting, and it decides whether you talk with an AI bot or sit there reading one. Beginners hit this alongside a handful of other patterns worth unlearning early.
