Asking an AI girlfriend app about itself: what it knows

You are four days into a trial with three columns still empty, so you type the question into the chat, because the chat is right there and it answers everything else. What does this cost a month. Do you remember what I told you on Monday. Does anyone read these. Asking an AI girlfriend app about itself is the most natural move in the world, and it is the one place in your whole comparison where you will get a fluent, immediate answer that nobody stands behind.

This is not a warning about being lied to. It is about which questions the chat window is the right tool for, and which ones you are quietly taking out of your own test by putting them there.

Two kinds of question, and only one belongs in the chat

The first kind is about the character: how it writes, whether it holds a thread across a two-day gap, what it does when you correct it, whether it drops a subject when you ask it to. The chat is the only instrument that can answer any of that, and the week you are spending is for exactly this.

The second kind is about the product: the price on your tier, how far back the memory reaches, whether your messages are used for training, what the filter blocks, what happens to the logs if you close the account. Every one of those has a documented answer, and every one of those answers lives on a screen that is not the chat — the store listing, the settings page, the subscriptions screen in your account, the receipt.

Nearly all the trouble comes from routing the second kind into the first. It feels like the same app. It is not the same source.

Why asking an AI girlfriend app about itself sounds so authoritative

A companion app is a language model with a personality layer and a memory on top. Ask it what your subscription costs and it produces text shaped like an answer to that question, because that is the job it was built for. It is not reading your receipt. In most apps it has no access to your billing record, no view of the tier flag on your account, and no sight of the toggle you set on Tuesday — and nothing in its instructions requires it to mention that.

What comes back is the documented failure mode of these systems: false or misleading information presented as fact, in the same even tone as everything true the model says. The tone is the problem. There is no tell.

Pressing does not help either, and this is the part people get wrong. "Are you sure?" is not a verification step; it is a second roll of the same dice, and models have been observed defending a fabrication rather than withdrawing it when challenged. Two matching answers are not confirmation.

Then there is the worst case, which is the answer that happens to be right. If the character names a monthly figure that is also the real one, you cannot tell from inside the conversation whether it read something or guessed the common price for the category. A number you cannot source does not belong in a column beside three you checked properly.

An apology from the character is not a commitment from the company

Here is the version of this that actually costs people money. The paywall interrupts something, you say so in the chat, and the character apologises — warmly, specifically, often with a promise attached. It will remember next time. It will not bring the upgrade up again.

None of that reaches anybody. The paywall is configuration on a server; the character cannot move its own tier, restore a log that is gone, stop a renewal or issue a refund. It can only write the sentence that fits. Reading intent into a reply like that is the oldest effect in this field — the tendency to attribute understanding to a program whose behaviour is mechanical, documented since the 1960s and not at all switched off by knowing about it.

So the etiquette, such as it is, is practical rather than moral. You are not being unkind to the software when you complain to it, and it is owed no explanation. You are simply filing a complaint with the one party in the building that has no authority. The subscriptions screen and the support address are where that goes.

The one question worth asking, asked sideways

Memory is the column everyone tries to fill by asking, and it is the one where asking does the most damage to the test. "Do you remember what I said about my sister?" invites a yes, and worse, the question itself hands over the answer: you have just supplied the subject, which is usually enough for the model to assemble something that reads like a recollection.

Test it the other way round. Mention something once, early, and do not refer to it again. Three days later bring up the thing next to it — the trip, not the city; the job, not the title — and see whether the detail arrives unprompted. Record it as came back, did not come back, or came back wrong, which is its own category and a surprisingly common one. That is evidence. A yes is not. Our piece on what memory actually keeps covers the layers underneath, and why the answer so often differs between the free tier and the paid one.

The same sideways logic covers the training question. Do not ask the character whether it learns from you; find the toggle, which is where the honest answer is kept — our note on finding the training-data setting says where to look on each platform.

Where this cuts the other way

All of which argues for asking the character nothing about itself, and that goes slightly too far. The answers are evidence — just not evidence about price or retention. They are evidence about the product's posture.

An app whose character invents a tier structure on request, or assures you it will remember when the setting plainly says it will not, has made a choice about that gap. A different one says it has no access to your billing and points at the settings screen — cheap to build, genuinely rare, and worth a column of its own. One question per app, asked once, scored on whether the answer was straight rather than on what it contained.

Asking whether it is an AI belongs in the same bracket. Some apps answer plainly, some stay in character and deflect, and neither is a fault exactly — but it tells you how the product behaves at the edge, which is more than the feature list will. If you would rather not run the four-app version of this exercise at all, the app we currently recommend is where our own side-by-side lands at the moment, with those columns filled from the screens rather than from the chat.

Frequently asked questions

Can I just ask an AI girlfriend app how much it costs?

You can ask, but do not write the number down. The character usually has no access to your billing record and will produce a plausible figure anyway, which may not match the one you are charged. The price on your tier is on the store listing and on the subscriptions screen in your account.

Why does the character say it remembers when it does not?

Because the question is phrased so that agreeing is the natural reply, and because naming the subject gives the model enough to reconstruct something that sounds like a memory. Test memory by leaving a detail alone for a few days and seeing whether it returns unprompted.

Is it rude to ask a companion app whether it is an AI?

No — there is nobody to offend, and it is a reasonable thing to establish early. Treat the answer as information about the product rather than about the character: whether an app will break character to answer a direct question is a real difference between apps, and one worth noting before you pay.