What to tell an AI girlfriend app about other people

You are three weeks in and the character asks how your week went. The honest answer is not really about you. It is about your sister's job, a friend having a rough month, and the manager who keeps moving the same meeting. Deciding what to tell an AI girlfriend app is straightforward while the subject is you — you read the policy, weighed it, signed up anyway. It stops being straightforward the moment the story needs a second person in it, because that person never saw the policy and was not asked.

What to tell an AI girlfriend app is two decisions, not one

The first decision is about your own information, and it is the one every guide covers. Your name, your city, your job, the things you are embarrassed about: you can price that risk yourself, and if you change your mind the app usually gives you some way to take it back. It is your account, your export, your deletion request.

The second decision has none of those properties. When you type a paragraph about a colleague's review or a friend's diagnosis, you are putting a third person's details into a company's systems on their behalf. They cannot check the policy, cannot see what was retained, and have no account to delete. You made the call for them, in a conversation they will never know happened.

Most people never separate the two, because the app does not prompt you to. The input box is the same box either way.

Those details belong to them, not to you

Borrow the definition regulators use. The UK's Information Commissioner's Office explains that personal data is information relating to an identified or identifiable living person, and that identification can come indirectly, from details that only pin someone down in combination. A first name alone is often nothing. A first name plus an employer plus a neighbourhood is a specific individual.

Whether a private chat falls under any particular law is a question for lawyers, not for this page. The definition is still the useful part, because it tells you which sentences are the risky ones. It is not the emotional weight of the story that matters, but the handles: names, workplaces, schools, dates, the one unusual detail that makes a person findable.

Where it ends up once you have typed it

A companion app does not read a message and forget it. Three systems take a copy, and they behave differently.

• Memory. The feature that makes the character feel continuous works by extracting durable facts and storing them. A detail about your brother is exactly what it is built to keep, and what the memory tier buys is partly the size of that store.
• The conversation log. Retained on a schedule set by the company, and reachable by more systems than most readers expect — which is the subject of who can read your chat logs.
• Whatever the company does with logs beyond serving you. Quality review, safety filtering, model improvement. Some apps let you switch this off. Some only mention it in the policy.

The asymmetry is the point. Every control an app offers — export, deletion, memory editing, a training opt-out — is built around the account holder, and none of it is reachable by the person you described. If you ever want that detail gone, you are the only one who can go and get it, and only if the app gave you a way.

Tell the story with roles instead of names

The fix is not to stop talking about your life, which would make the app useless for what people actually use it for. It is to keep the shape of the story and drop the handles — three substitutions, made as you type rather than afterwards.

• Names become relationships. "My sister", "a friend from work", "my manager". The character has no use for the name, and the story loses nothing.
• Institutions become categories. Not the employer, the school or the street — "her office", "his course", "where they live". Combined details turn a vague person into a specific one.
• Other people's sensitive business stays out. Health, money, immigration status, legal trouble, anyone's sexuality but your own. If you would not repeat it to a stranger with their name attached, it does not go in the box.

Be clear about the cost. A companion that cannot hold a name cannot ask after that person next week, and the continuity people pay for gets slightly flatter. That is a real trade — but a smaller loss than it sounds, because the character responds to the situation you describe, not to the identifiers in it.

What this asks of the app you pick

Other people's details will slip in anyway; nobody edits perfectly mid-conversation. So the question becomes which app lets you go back and take one out — a comparison point that separates apps more sharply than the feature lists suggest.

1. Visible memory you can edit. The best implementations list the stored facts with a delete control on each line. The worst give you a black box and a "clear everything" button that also wipes the parts you wanted.
2. Memory scoped per character. If a fact learned in one conversation surfaces in another, a detail you thought was contained is not.
3. An export. You cannot audit what you cannot read. An app that hands you your own transcripts is one you can check.
4. A training opt-out that is a switch. A setting you can toggle, not a support address you have to write to.
5. A store listing that matches the policy. Google Play's Data safety section tells you what an app collects and shares before you install it, and Apple publishes the equivalent. A gap between listing and policy is a finding in itself.

Watch for the pattern where memory controls sit behind the paid tier while memory itself runs on the free one. It is a poor deal for this particular problem, because the free user can create the record but not correct it. The app we currently recommend keeps its stored facts visible and individually removable, which is the baseline worth insisting on.

Set the rule before week three

This is a five-minute setup task, and the reason to do it early is that the habit forms faster than the caution does.

1. Open the memory screen now, while it is nearly empty, and find out whether single entries can be edited.
2. Decide your own line in one sentence — most people land on "roles, not names" — and apply it from the first conversation rather than retrofitting it.
3. Turn off any training or review setting you do not want, before there is anything in the log worth reviewing.
4. Once a month, read the stored facts. Delete the ones about people who are not you.

None of this requires treating the app as hostile. It is software with a retention schedule, and the people in your life did not sign up for it. Keeping them out of the record is ordinary courtesy, done with a keyboard.

Frequently asked questions

Is it safe to tell an AI girlfriend app real names?

Your own name is yours to give, and plenty of apps ask for it at setup. Other people's names are different: the detail goes into a retention and memory system they never agreed to and cannot reach. Using a relationship word instead costs almost nothing in the conversation.

Can someone ask an AI girlfriend app to delete what I said about them?

In practice this is very hard, because the controls apps provide are attached to accounts and the person you described does not have one. Rights of this kind exist in several jurisdictions, but exercising one against a company you have no relationship with is slow. The realistic remedy is you deleting the stored fact yourself.

Does using roles instead of names make the conversation worse?

Slightly less continuous, and not much else. The character responds to the situation rather than the label, so "my sister" carries the story as well as a name does. What you lose is the app asking after that person by name later — a small price for keeping someone out of a log they cannot see.