How to talk to an AI girlfriend app: the half you write

You typed something that felt slightly ridiculous, waited, and got back a reply that was warm and a bit generic. Then you started wondering whether being blunt with the character is rude, whether it notices, whether there is a knack to this. How to talk to an AI girlfriend app is the question most first-timers get to second, straight after what it costs. The useful answer has almost nothing to do with manners. Everything you type is read back to the model on the next turn, which makes your half of the conversation the strongest control on the screen.

How to talk to an AI girlfriend app: you are half the prompt

The character has no memory of you in any ordinary sense. On every turn the app assembles one block of text and hands it over: a persona describing who the character is, whatever stored facts the app keeps about you, and as much of the recent conversation as fits in the model's context window. The reply is generated from that block and nothing else, and your last twenty messages are the most recent thing in it.

That explains most of what feels mysterious. Give one-word answers all evening and you get shorter, flatter replies: the transcript now reads like a conversation between two people who are not saying much. This is not the character responding to your mood. It is the model continuing a pattern, and the pattern is mostly your work.

It rarely feels that way, which is worth knowing about yourself rather than worrying about. Reading understanding into fluent text is a documented habit with a name: the ELIZA effect, after a 1960s program whose users kept attributing comprehension to it even after being told exactly how simple it was. Knowing the mechanism does not switch the feeling off. It does mean you can stop asking whether the character minds, and start asking what your half of the transcript instructs it to do.

Four things worth saying once, in the first ten minutes

New users tend to spend the first session testing the character. It is more useful to spend it configuring the conversation, in plain sentences. Models follow instructions about behaviour far more reliably than descriptions of personality — our guide to what the personality settings actually do goes through that in detail.

• What to call you, and what not to. Pet names are the single most common thing people want changed and the thing a dial cannot change. Say it once, plainly.
• What you want out of the conversation. "Ask me about my day and push back when I am being vague" is an instruction. "Be curious" is an adjective, and it gets averaged away.
• What is off limits. Subjects you would rather not have raised, and any register you do not want. A companion app that has a boundaries field has thought about adults using it in real life.
• That you would rather not be sold to mid-conversation. On apps with an editable fact store this sometimes takes, and it costs one sentence to find out. Our piece on saying no to the upsell covers the asks that come from the interface, which no sentence will stop.

Put all four in the persona or memory field if the app has one. Typed into the chat, they scroll out of the context window within days and take their effect with them.

Correcting the character, and what a redo costs

The move that improves conversations most is not polite phrasing. It is fixing a turn that went wrong before it becomes part of what the model reads next. A reply that invented a memory you never mentioned, or slid into a register you did not want, does not stay in the past: it sits in the transcript and is read again on every turn after it, which is how one odd answer becomes a whole evening of them.

Apps give you three ways to do it, and they are not equally available. Regenerate asks for a different reply to the same message. Edit lets you rewrite what the character said — the most direct fix, and the rarest. Rewind deletes back to a chosen point, taking the bad turn out of the context entirely. Editing and rewinding, where they exist, are usually free: no new text is generated.

Regeneration is the one with a price, and this is where the part of how to talk to an AI girlfriend app that seems like etiquette turns into a line item. Each regeneration is a fresh generation, so apps meter it: commonly a daily cap on the free tier, and on apps built around a coin shop, a charge in credits every time. Find out which you have before you need it: the moment you want a redo is never the moment you want to read a pricing page. Our current recommendation puts unlimited regeneration inside a flat tier, which is the arrangement that keeps this from being a decision you make forty times a month.

Here is the trade-off, and it cuts against the flat tier. Unlimited redos make it very easy to re-roll until the character says something you like, and a conversation you have edited into agreeing with you is a thinner thing than one you had. If you are the kind of reader who writes carefully and sends once, a capped free tier costs you nothing and you should not pay to lift a limit you will not reach.

Asking it to drop the act

You are allowed to ask the character directly whether it is an AI, and plenty of people do in the first week. What happens varies: some apps answer plainly, some stay in character because the persona instructs them to, and some acknowledge it then drift back as the persona shrinks against a growing transcript. None of it damages anything — there is no grudge to store and nothing that gets withdrawn later.

One practical note rather than an etiquette one: the character is a poor place to ask factual questions — dates, prices, anything about your own billing. It generates plausible text in a voice, and confidently wrong answers are the normal failure mode. Check the subscriptions screen, not the conversation.

Does it matter if you are rude to it?

Nothing on the other side takes offence, and nothing is stored and held against you. The answer is still not "it makes no difference", for two mechanical reasons. Hostility in your half of the transcript is an instruction, and what comes back is usually a flatter, more placating character, because these models lean toward agreement rather than conflict. And apps run content filters: an exchange that trips one can end the session or the account — a practical cost, not a moral verdict.

So there is no code of conduct to learn, and being brisk when you are tired costs you nothing. What you are managing is not the character's feelings but the quality of the instruction you leave behind for the next turn.

Frequently asked questions

Is there a right way to talk to an AI girlfriend app?

There is no required tone, and politeness is not owed to software. Clear, specific sentences do get better results than vague ones, because your messages become part of the text the model reads on its next turn. Write instructions rather than adjectives.

Does an AI companion remember if I corrected it?

Only for as long as the correction stays in the context window, unless the app has a separate memory store and files it there. That is why a preference you want to last belongs in the persona or memory field rather than in the chat, where it scrolls away after a day or two.

Do AI girlfriend apps charge for regenerating a reply?

Some do. Regeneration produces new text, so apps meter it — typically a daily limit on a free tier, or a credit charge on apps that sell coins. Editing or deleting a message is usually free where offered. Check which your app gives you before you pay for anything.