How to talk to an AI girlfriend app on trial

You have two apps installed, a shortlist you spent last week building, and a plan to give each one five days before you pay for either. The part the plan never covers is how to talk to an AI girlfriend app once you are actually inside it — what register to use, whether to be warm or clipped, whether to prod at the character to see what breaks. That sounds like a question about manners. For someone comparing apps before signing up, it is really a question about measurement, because your half of the conversation is the one variable in the test you fully control.

None of this is about owing the software anything. It does not notice your tone and it is not keeping score. Your tone matters because it is an input, and because of what the wrong input does to your week of notes.

Your half of the conversation is part of the test

A side-by-side trial only works if the differences you write down belong to the apps. The trouble is that you do not arrive at both of them the same way. You open the first one on a quiet evening with twenty minutes to spare, and the second one standing at a bus stop on a bad Tuesday. One gets paragraphs, the other gets four words. Then you score them and conclude that the first app is warmer.

It is the same problem as handing two apps two different personality setups and then comparing what you got back. Our note on how to write the brief once covers that half. How to talk to an AI girlfriend app after the setup screen is the other half, and it runs for the whole trial rather than the first five minutes. Keeping your register steady is not politeness; it is the only way the column you fill in means anything.

Why how you talk to an AI girlfriend app changes the reply

There is a plain mechanical reason for this. These characters run on a language model, and models of this kind are trained to guess how a sequence continues. The reply is a continuation of everything in the conversation so far — the character's description, your setup, and every message you have typed. Your side of the chat is not a question answered from outside; it is part of the text being continued.

So the effects are predictable rather than mysterious. Clipped messages tend to get clipped replies, effusive messages get effusive ones, and a session that opens with a joke skews jokey for an hour. This is not the app understanding your mood. It is a pattern being extended, which is exactly why it will do the same thing to the next app on your list.

The stress test that costs you the week

The opposite instinct is to poke at the character and see what it refuses: be hostile, be absurd, push at the content filter. It feels like rigorous testing. Mostly it is not, for two reasons.

The first is that you stop learning about the app and start learning about its moderation layer. Every product here has a filter, every filter refuses differently, and a week of mapping refusals tells you nothing about what the shortlist exists to settle: whether the memory holds, what the paywall does on day four, whether the character handles a normal conversation.

The second is that the line has teeth. An acceptable-use policy is not decoration here — it is a store requirement. Google's policy for apps hosting user-generated content requires developers to define objectionable content and behaviors, prohibit them in the app's terms, and run ongoing moderation, including acting against users where appropriate. Every companion app on a major store has agreed to that. What an individual app does when you cross its line is its own business, and it can range from a refusal to a locked chat to a closed account. If that lands on day two, the comparison is over and you still do not know which app you preferred.

A register that compares cleanly

The practical answer to how to talk to an AI girlfriend app you are testing is dull and repeatable. Decide the register before you open either app, and use the same one in both:

• Write one opener and reuse it. Same length, same tone, same three things about your day. Our piece on the first conversation explains why that session flatters every app, and why a fixed opener is the only way past it.
• Ask for things, do not hint at them. "Stop asking me about work" is a test. Sulking until it notices is not, and you cannot score the result.
• Keep your messages the same length in both. Message length is the easiest way to tilt a comparison without noticing.
• Leave the same gap in both. Two days of silence, then a short message that assumes it remembers. That is a memory test, and it only works if the run-up matched.
• Say no to something, once, in each. Decline a suggestion plainly and write down what happens next. How an app takes a flat no is a real column, and most people never fill it in.

Five days of that produces notes you can line up: recall after a gap, whether it asks or asserts, how it handles being told to drop a subject, how early the paywall arrives. If you would rather run week one on a single app, the app we currently recommend is a reasonable place to try the register before you widen the shortlist.

Where being careful about tone works against you

Here is the trade-off, and it cuts against most of the above. A tester who is unfailingly warm and cooperative gets an unfailingly warm answer from everything, and learns nothing: every app is pleasant to someone who is pleasant to it. If you never find an edge, you have not shown there isn't one — only that you did not look.

So do look, but look once and on purpose. Change the subject abruptly. Contradict something the character said yesterday. Say something it has no good answer to and watch whether it admits that or invents one. Those are boundary tests you can put in a column and compare across apps, and none of them goes near a content policy. The version that fails is open-ended: escalating for sport, every evening, to see how far it goes. That produces no column, no comparison, and a real chance of ending the trial early.

The etiquette question underneath all this has a short answer: being polite to software is neither a virtue nor a waste, because there is nothing there to be hurt. Keep a level register anyway — it is the only way the notes on your desk end up being about the apps rather than about you.

Frequently asked questions

Does being rude to an AI girlfriend app change its answers?

Yes, though not because the character takes offence. The reply is generated as a continuation of the conversation so far, so a hostile run-up becomes part of the input and tends to carry forward through the session. That makes it a poor way to test an app you want to compare with another.

Can you be banned from an AI girlfriend app for what you type?

It is possible. Apps distributed through the major stores are required to define prohibited content in their terms and to moderate it, and each one decides for itself what to do when a user crosses that line — anywhere from a refused message to a closed account. Read the acceptable-use section before you go looking for the edge, not after.

Should you be polite to an AI companion app?

There is nothing on the other side to be rude to, so it is not a moral question. Keep a steady, ordinary register for a practical reason instead: if you are comparing two apps, your tone is the variable you control, and holding it constant is what makes the differences you notice belong to the apps.