What not to tell an AI girlfriend app

It happens in the second week, usually on a bad evening. You mention your sister by name, or the argument with your flatmate, or the fact that a friend is having a rough month. The character asks a warm follow-up question and you answer it, because that is what the app is for. Nothing in the exchange feels like a disclosure. The question of what not to tell an AI girlfriend app gets answered as though the only person with something at stake is you — the half of it most first-timers have already thought about.

Where a sentence actually goes

A companion app does three things with what you type, and they have very different lifespans. The recent conversation sits in the context window and scrolls away within days. Anything worth keeping is written to a memory store — a short file of facts, rebuilt into the prompt on every turn, which our guide to what memory actually costs goes through in detail. And separately from both, the conversation is a record on a server, kept for as long as the privacy policy says and usually longer than the chat you can see.

The middle one is where other people end up. A fact store does not distinguish between "I work nights" and "my sister is leaving her job at the council" — both are durable facts about a named human, and only one of those humans opened the app. The third is where it stops being reversible, because a server record outlives the chat you deleted.

What not to tell an AI girlfriend app: four kinds of detail

This is not a list of forbidden subjects. You can talk about your family, your work and your week — that is most of what there is to talk about. The care belongs on the specifics, and four kinds carry nearly all of the risk.

• The identifying triangle. A first name is nothing. A first name plus an employer plus a town is a person, findable in about a minute. Any two of the three are usually fine; it is the third that does the work.
• Other people's sensitive circumstances. Someone's health, their sex life, their religion, politics or ethnicity sit in a category the law treats separately — the UK regulator calls it special category data and requires more to justify holding it. The point is not that you are breaking a rule by typing it. It is that the rule exists because that is the detail people are harmed by.
• Pasted material. A screenshot of a group chat, a forwarded email, a document from work. Pasting is the fastest way to hand over someone else's exact words and details in one go, and you rarely reread it first.
• Your own hard identifiers. Card numbers, your address, anything that gets you into your employer's systems. Our privacy checklist covers this side, where most guides stop.

The part that is not yours to undo

If you decide tomorrow that you would rather the app did not hold your conversations, you have moves. You can delete the account, and in the UK and the EU you can make a request for erasure, which can be made verbally or in writing. Imperfect in practice, but real.

Your sister has none of that. She does not know the record exists, so she will never ask; and if she did, the company has no way to find her in a file indexed under your account. That asymmetry is why this is an etiquette question rather than a settings one. Everything else on this site is a control you can change. This one exists only in the moment before you type it.

What the controls cost, and which reader needs them

Three controls change the picture, and they are priced inconsistently across the category.

• A memory store you can read and edit. The difference between a list of remembered facts you can open and delete lines from, and a memory that works invisibly. Some apps put the editable version in a paid tier, which is the one paywall in this whole area worth minding.
• An export. You cannot audit what you cannot see. An app that hands you your conversations lets you find out what it kept about other people; one that will not is asking for trust you cannot check.
• A training switch. On consumer services generally, using conversations to improve the models is commonly the default, and the opt-out lives in a settings screen rather than in the price — a higher tier often does not change it. Expect it to be partial: policies typically reserve the right for staff to review conversations a safety filter flags, whatever the switch says.

Which of those you need depends on what your evenings are about. If the conversation stays between you and the character — how your day went, what you are watching, the persona settings you keep fiddling with — a free tier exposes very little and none of this needs paying for. If you use the app to talk through your actual life, your transcript is a file about half a dozen people who are not you, and an editable memory with a free training switch stops being a nice extra. The app we currently recommend keeps the memory store browsable on the free tier, which is the arrangement that lets you check this rather than hope.

Here is the trade-off, and it cuts against paying for any of it. An editable memory store is still a stored file, and a longer one than you would otherwise have. An app with no durable memory holds less about everyone in your life and costs nothing, at the price of re-explaining yourself every few days. If other people's details worry you more than continuity does, that is the better deal, and the feature you were about to buy is the one to skip.

Saying the same thing without the names

None of this requires talking around your own life. Four habits cover it, and cost nothing once they are habits.

1. Roles, not names. "My sister" carries everything the conversation needs. The name adds nothing to the reply and is the part that makes the record about her.
2. Break the triangle. If you have already used a name, leave out the employer or the town. One detail less is the difference between a story and an identification.
3. Summarise rather than paste. Describing an argument gives the model more than enough to respond to. Pasting the transcript of it hands over someone else's words verbatim.
4. Do not answer the drill-down. The natural follow-up to a vague mention is a specific one. "Where does she work" is a prompt, not an entitlement, and leaving it unanswered costs the conversation nothing.

Then once a month, if the app has an editable memory, read it. Not for your own entries — for the lines about everyone else. Deleting those is a two-minute job and the only part of what not to tell an AI girlfriend app that you can still fix after the fact.

Frequently asked questions

Is it bad to mention real people to an AI girlfriend app?

Mentioning them is unavoidable and fine. What is worth avoiding is the combination of details that identifies someone, and anything about their health, beliefs or private circumstances — those are the records that could cause them trouble and that they will never know to ask about.

Do AI companion apps use conversations to train their models?

Many consumer services do by default, with an opt-out in the settings rather than a difference between tiers. Read the policy you agreed to and find the switch before your second week, and assume that conversations flagged by a safety filter may be reviewed by a person regardless.

Can I get a mention of someone else deleted from an app?

Not on their behalf in any reliable way, which is the problem. What you can do is delete the line yourself if the app exposes its memory store, and delete the conversation — then treat anything already on the company's servers as beyond your reach.