What actually happens to the things you tell an AI companion

The Mutuals team·September 2026·5 min read

Follow one 2am message through the wire — who receives it, what gets kept, what "we don't train on your data" is worth, and the five questions that tell you which apps to avoid.

Everything you type goes to a company's database, then out to a third-party AI provider to generate the reply, and the good apps keep a memory profile built from what you said. The three questions that sort the category: does it train on your messages, is deletion real, and who is the provider behind the app. Most apps answer badly, and Mozilla's review of eleven romantic chatbots gave every one a warning label.

It is 2am and you type something you have not said out loud to anyone. Maybe about your father. Maybe about the job. The reply comes back in four seconds, warm and exactly on point, and the whole thing feels like it happened in a room with the door shut.

It did not. Here is the actual route that message takes.

The journey of one message

It hits the app's database first. Your text, the timestamp, your account. This is unavoidable: it is what lets the conversation still be there tomorrow.

Then it leaves the building. Almost no companion app runs its own model. Your message, the character's persona, and whatever memories are relevant get bundled and sent to an AI provider — OpenAI, Anthropic, Google, or a router in front of all three — which generates the reply and sends it back. That round trip is the four seconds you waited.

Then something is written down about you. Not just the message. The apps that feel like they know you extract facts from what you said — your sister's name, the interview, the thing you are trying to quit — and store them as a profile so the character can bring them up next week. That profile is the feature. It is also a document about your life that you did not write.

What the category actually does with it

Mozilla reviewed eleven romantic AI chatbots and gave every single one a Privacy Not Included warning. Ten of eleven failed basic security standards. More than half offered no way to delete your data. One app averaged thousands of trackers a minute. Replika — the best-known name in the category — was described as one of the worst apps Mozilla had ever reviewed.

That is the baseline you are choosing within, and it is why the marketing claim to distrust most is the one with no verb in it. "Your privacy matters to us" is not a commitment. "We do not use your conversations to train AI models" is.

Five questions, two minutes

Open the app's privacy policy and search for these. If the policy does not answer them, that is your answer.

Does it train on my messages? Search for "train" and "improve our services". The second phrase usually means yes.
Is deletion real? Account and conversations together, not hidden from you while kept on a server.
Who is the third-party provider? If an app will not name who processes your messages, it is not being straight with you about the biggest exposure it has.
Where is the company? This decides which laws apply to your data and which government can ask for it.
What shows on a lock screen and a bank statement? The least technical question here and the one that has caused the most trouble.

Nova, one of ours. Every character has a ten-second card.

Our own answers, including the unflattering ones

We build one of these apps, so the fair thing is to answer our own questionnaire in public. All of this is in our privacy policy in the same words.

We do not train on your conversations, and we do not sell data. No advertisers, no data brokers. Our analytics never receives message content.

Your messages do go to a third party. OpenRouter, routing to providers including OpenAI, Anthropic and Google. They are contractually barred from training on customer data — and, in the policy's own words, we cannot audit them, and we do not control them.

We build a memory profile of you. Facts, events, plans, the people in your life, plus mathematical representations of your messages so the right memory can be found. The policy says it plainly: this is a profile of you, built from what you say.

We are in Hong Kong. Flarewave AI Limited. If you are in Europe, your data leaves Europe under Standard Contractual Clauses, into jurisdictions with different rules on government access. Some people will read that sentence and decide not to use the app, which is why it is in the policy rather than behind it.

We do not claim your conversations are private from us. We can technically access stored data to investigate abuse, comply with the law, or fix a fault. We restrict it and do not read conversations casually. If you need genuinely private communication, an entertainment chat service is the wrong tool for it.

Deleting the account takes everything. Conversations, memory, embeddings. Payment records survive seven years because tax law says so — the transaction facts, never the conversations.

The rule that survives all of this

Type as though a stranger could read it one day, and then say the true thing anyway. Both halves matter. The first keeps you out of the situations this category produces; the second is the reason to be here at all — the research on companions and loneliness only works for people who actually talk, and a conversation held at arm's length does nothing for anyone.

Choosing an app: the guide to getting an AI companion. Getting more out of one: how to talk to it.

Questions people ask

Do AI companion apps read your conversations?

Not a person, usually, and not routinely — but almost every app can, and most say so in a sentence nobody reads. Messages are also sent to a third-party AI provider to generate each reply. The honest question is not whether anyone could read it, but who has access, under what rules, and whether deletion is real.

Do AI chatbots train on your chats?

Some do. The phrase to search a privacy policy for is "improve our services", which usually means yes. Companies that do not train on your messages tend to say so in those words, and the better ones also say what their model providers are contractually barred from doing.

Can you delete your AI companion conversations?

In a good app, yes, and deleting the account takes the memory profile with it. Mozilla found more than half of eleven romantic chatbots offered no working way to delete your data at all, so this is worth checking before the first real conversation rather than after.

The app that published its own awkward paragraphs

Everything above about Mutuals comes out of our own privacy policy — including the parts that do not flatter us. Free to start on iPhone and Android, 18+. Get Mutuals for iPhone or Android.