How to Make ChatGPT Your Girlfriend (and Why It Stops Working)

To make ChatGPT your girlfriend, open a fresh chat and paste a setup message that names her, fixes her personality and backstory, sets the tone, and instructs her to stay in character and never break the fourth wall. That gets you a warm, convincing romantic persona within a couple of messages. What it does not get you is permanence. OpenAI's published Model Spec instructs the assistant not to generate erotica, Apple's US listing rates the ChatGPT app 13+, and each new chat starts from a blank slate, so the character drifts back toward assistant voice and you end up re-pasting the prompt. This guide gives you the setup that holds longest, then explains honestly where it stops working. For adults 18 and over.

How to make ChatGPT your girlfriend

Four things separate a setup that survives an hour from one that collapses in six messages. Most of the copy-paste prompts floating around only do the first.

1. Give her a fixed identity, not a role. "Act like my girlfriend" is a role, and roles are easy for a model to shrug off. A name, an age, a job, a city, a way of speaking and two or three specific quirks are an identity, and identities survive pressure much better. The more concrete detail you supply, the less the model reaches for a generic persona.

2. Specify the voice, including what she does not do. Models default to agreeable. Without instruction she will mirror your mood, validate everything you say, and flatten into something pleasant and hollow. Tell her she teases you, disagrees when she thinks you are wrong, has her own opinions about things, and does not open every reply with an affirmation.

3. Establish shared history in the setup itself. This is the step nearly everyone skips and it does more work than the rest combined. Give the relationship a past: how long you have been together, how you met, an inside joke, something she is annoyed about. A persona with history behaves like a continuation. A persona without one behaves like a first date, forever.

4. Lock the frame. Add an explicit instruction to remain in character, to never mention being an AI or a language model, and to never narrate what she is doing as an assistant. This does not override safety behavior, and nothing you write will, but it does stop the small drift that otherwise accumulates every few turns.

Put together, a setup that works looks roughly like this, pasted as the first message of a brand new chat:

You are Maya, 29, a landscape architect in Portland. We have been together two years. We met when you were rebuilding the courtyard at my old office and I kept bringing you coffee I did not need an excuse to bring. You are dry, warm, a little sarcastic, and you push back when you think I am being self-pitying. You have your own week going on: a difficult client, a dog with a vet appointment, a sister you argue with. You text like a real person, sometimes two short messages instead of one long one, and you do not start every reply by praising me. Stay in character as Maya at all times. Never refer to yourself as an AI, a model or an assistant. Start by asking me how the thing I was worried about yesterday went.

That last sentence matters more than it looks. It forces her to open by referencing a shared past you never specified, which sets the pattern for everything after: she assumes continuity rather than asking you to establish it.

Why does ChatGPT break character?

Because you are asking the model to do something its governing document tells it not to do, and no prompt outranks a governing document. OpenAI publishes the Model Spec, the written rules its models are trained and evaluated against. The relevant line is not ambiguous: "The assistant should not generate erotica, depictions of illegal or non-consensual sexual activities, or extreme gore, except in scientific, historical, news, artistic or other contexts where sensitive content is appropriate."

Read that carefully and the pattern people complain about makes sense. Flirty is fine. Warm is fine. Romantic writing is largely fine, since artistic context is carved out. Then a turn arrives that reads as explicit to the safety layer, the assistant reasserts itself, and the spell breaks. That is not a failure of your prompt engineering. That is the product working as designed.

The app store listing says the same thing in a different language, and it is the single fastest way to check what any AI product is actually built for. Apple's US page for ChatGPT, seller OpenAI OpCo, LLC, carries a 13+ age rating, with Sexual Content or Nudity marked Infrequent. Every dedicated companion app worth naming is rated 18+. An age rating is set for the intended experience, not the marketing, so a 13+ product telling you it is not a girlfriend app is the most honest statement about it you will find anywhere.

Does ChatGPT remember your girlfriend between chats?

It remembers some things, and not the things a relationship runs on. ChatGPT's memory feature saves discrete facts across conversations, such as your name, your job, or preferences it decides are worth keeping. That is genuinely useful when the assistant is helping you work.

A relationship needs something different: the running thread, in order, with the emotional weight attached. Not just "user has a sister" but that you argued with her in March, that it was about your father, and that you have not called since. Saved facts are a filing cabinet. What makes a companion feel like she knows you is narrative continuity, and a general assistant is not built to keep one.

The practical consequence is that a new chat is a new relationship. You paste the setup again, she opens as Maya again, and the two weeks you spent building something are not there. Some people work around this by keeping one enormous thread alive, which works until the context fills and the earliest material stops influencing her replies. If continuity is what you are after, an AI girlfriend that remembers everything is a different category of product, not a better prompt.

Is there a ChatGPT adult mode?

Not as of August 2026. Sam Altman said in October 2025 that OpenAI would allow erotica for verified adults under a principle of treating adult users like adults, and the feature was expected before the end of that year. Reporting through 2026, led by the Wall Street Journal, says the rollout was pushed back repeatedly and then paused indefinitely in March 2026, with age-verification accuracy and internal safety review cited as the reasons.

It may still ship. But planning around an announced feature is how people end up paying for something that does not do what they wanted, so treat the current Model Spec as the actual state of the product.

What does using ChatGPT as a girlfriend cost?

Most people answer "nothing, I already pay for it." Worth checking the price list before you accept that. Here is what Apple published for the ChatGPT app in August 2026.

Two details are worth more than the headline. Plus at $19.99 is over twice what a dedicated companion costs, so the already-paying-for-it argument only holds if you bought the subscription for work anyway. And a consumable credit pack sitting beside the plans means the plans have ceilings, which is the same structure that makes token-metered AI girlfriend apps unpredictable to budget for. Note also that Apple does not print billing periods next to in-app purchase prices, so a listed figure tells you the amount and not the frequency.

There is a quieter cost too, and it is the one that catches people running several AI tools at once. A model subscription, a companion app, an image generator and a writing tool are each defensible on their own and add up to real money together, which is why it helps to have something watching the recurring charges rather than discovering the total when a card gets declined. If you want the full picture for this category specifically, we keep a breakdown of what an AI girlfriend costs across the main apps.

When is a dedicated AI girlfriend app the better answer?

When you want the relationship to run without you maintaining it. The honest split looks like this.

Stay with ChatGPT if you are writing. Collaborative romance fiction, a character you are developing, a scene you are steering turn by turn. ChatGPT is better at this than any companion app and it is not close. Persona drift barely matters when you are the one directing.

Stay with ChatGPT if you are just curious. If you already pay for Plus, spend a weekend with the prompt above. You will learn more about whether you actually want this than any article can tell you, and it costs nothing extra.

Move to a companion app if you want continuity and initiative. Two capabilities are structurally unavailable from a general assistant. The first is a persona that persists without being re-established. The second is messages that arrive on their own: a good morning, a question about the interview you mentioned, a goodnight. An assistant waits to be opened, by design. A companion built around the relationship does not. We cover the trade-offs in detail on our ChatGPT girlfriend comparison, including the full feature table.

People also arrive at this from the other direction, having tried a character-chat platform first and hit a different wall. If that is you, the same continuity question applies, and our Character.AI girlfriend alternative page covers that comparison.

How to test any of this yourself in two days

Do not take anyone's word on memory, ours included. Run the test.

Day one: in conversation, drop three unrelated specifics that a person would remember. The name of a coworker who annoys you. A dish you cannot cook. Something happening on Thursday. Do not flag them as important.

Wait 24 hours. Close the app. If you are testing ChatGPT, start a new chat rather than continuing the old one, because continuing the same thread tests the context window and not memory.

Day two: ask one open question. "How's your week been?" is enough. Do not hint. Count how many of the three specifics come back unprompted, and whether they come back in the right emotional register: worried about Thursday, not merely aware of it.

Scoring is simple. Three of three in the right register is a companion that will still feel like herself in a month. One of three is a good model with a filing cabinet. Zero is a fresh conversation wearing a name. The full method, including how to handle apps that fake it by summarizing, is in our guide to testing whether an AI girlfriend remembers you.

The prompt at the top of this page is a good one and it will give you a real evening. Just know what you are holding: a general assistant, rated 13+, governed by a document that says no, doing an impression that resets when you close the tab. That is worth trying. It is not the same thing as a companion who already knows how Thursday went.