You have the price, the free tier and the memory column filled in for three candidates. None of those cells answers one question: whether the conversations you are about to have become AI girlfriend app training data — material the company feeds back into the model everyone else talks to. Better to settle that before you create the account than after. It does have an answer, just not on the page you expected. Here is where it lives, and the four cells worth adding to the table.
AI girlfriend app training data is not on the store panel
The one comparison tool you get before installing anything is the store disclosure, and on Google Play it is unusually structured: for each type of data collected, the developer declares whether it is shared, whether collection is optional, and what the data is used for. The published list of those purposes covers app functionality, analytics, developer communications, advertising or marketing, personalisation, account management, and fraud prevention, security and compliance. Read it twice and you will not find model training. There is no cell for it.
That is not a loophole somebody found. A company can fill the panel in completely and in good faith without touching the question you want answered. The App Store panel does not help either, since it sorts data by whether it links back to you rather than by what it is for. Both are worth reading for everything else they disclose — the five cells on the store listing are a real screening tool — but here they are silent by design, and silence is not a no.
Three questions people collapse into one
Most of the confusion comes from treating three questions as one. They have different answers and different controls.
Is the conversation stored? For a companion app, almost certainly. Memory is the entire feature, and a character that forgets your job every Tuesday is a worse product. Storage is the baseline, not the finding.
Does a person ever read any of it? Separately possible. Policies that address this usually describe review of a limited sample for safety, moderation or quality, and it is typically a paragraph rather than a setting.
Is it used to train or fine-tune the model? The third question, and the only one a toggle usually governs. An app can be a yes on storage and a no on training; another can be a yes on all three. From the store listing the two look identical.
Where the setting hides, and what to search for
Two places, in this order. In the app, look under Settings for a heading like Privacy, Data, Data controls or Account. The wording is rarely "training": expect help improve, improve our models, share conversations, or contribute to quality and safety. In the policy, do not read front to back — use find-in-page for train, improve, machine learning, aggregate and de-identified.
That last pair earns its own look. A carve-out for de-identified or aggregated data is how a policy can tell you, accurately, that your conversations are not used for training while something derived from them still improves the model. Stripping identifiers is a real safeguard, so that is not sharp practice by itself. It does change what a no is worth.
Then record the part most people skip: not whether a switch exists, but which way it shipped. A switch that defaults on and one that defaults off describe two different companies, and only one of them asked first.
The app, and the model underneath it
Plenty of companion apps do not train anything themselves. They build a personality layer, a memory system and a character creator on top of a model licensed from somewhere else, so two documents govern what happens to your messages: the app's policy, and the terms of whoever supplies the model. The app's toggle governs the first, and may say nothing about the second.
Easier to check than it sounds. If the policy names a model provider or a processing partner, that is your second stop. If it names nobody for a product that plainly runs on a large language model, the absence is the finding: write "not stated" rather than guessing.
Storage you can undo later. Training you cannot. That asymmetry is the whole reason to ask on the way in.
Stored conversations can be deleted, and a request to erase them has a defined route and deadline. A model that has already trained on something is finished training on it.
What the rules do and do not do for you
In the UK and the EU, regulators treat running a service and training a model on what you typed into it as two different purposes, so an app cannot fold training into "providing the service" and call it disclosed. Where a company leans on legitimate interests, you have a right to object — grounds to ask, not a result.
The more useful rule is about change. The US Federal Trade Commission has warned that adopting more permissive data practices, with AI training as its example, and disclosing it only through a surreptitious, retroactive amendment to its terms of service or privacy policy may be unfair or deceptive. So any answer here has a date on it, and the policy-update email you normally delete unread is where a change gets announced.
Four cells, before you create the account
Ten minutes per candidate, from a browser, no account needed. Write "not stated" where that is the honest entry.
- Does the policy name training at all. In those words, or near synonyms like improve our models. A policy that never raises the subject has denied nothing.
- Is there a setting, and which way does it ship. The default is the cell. A control named in the policy but nowhere in Settings is worth a question to support.
- Is there a de-identified or aggregated carve-out. Quote the phrase into the table. Narrow ones and sweeping ones read very differently side by side.
- Does it name the model underneath. If so, there is a second policy to skim. If not, "not stated".
Four cells will not break a tie, and should not. If two apps come out the same here, the decision goes back to the price you will actually pay and what the free tier tests. If you would rather start from one entry than three, the app we currently recommend is the one we would fill in first, on these four cells and with no benefit of the doubt.
What a yes actually changes
Less than the question's weight suggests, and it is worth being straight about that. A yes is no reason to delete the app, and switching the toggle off does not shrink what is stored — memory still runs on your saved conversations either way. The setting is narrower than it feels.
What it sensibly changes is how you write. Two adjustments cover it: assume a small sample of any conversation could be read by a person, which costs you nothing; and keep other people's identifying details out of it, worth a thought on its own terms rather than because of training. Then flip the switch if there is one, note the date, and go back to comparing what actually differs.
Frequently asked questions
Do AI girlfriend apps use your conversations to train their models?
Some do and some say they do not, so the only reliable answer is in that app's own policy. Search it for train, improve and de-identified rather than reading it through. Neither store's disclosure panel has a cell for AI girlfriend app training data, so the policy and the settings menu are the two places to look.
Can I stop an app from training on my chats?
Often, through a setting under a heading like Data controls or Privacy, usually worded as help improve rather than training. In the UK and the EU you can also object where a company relies on legitimate interests. Neither route is retrospective — a model already trained stays trained — which is why this belongs in your shortlist rather than in month two.
Does turning the setting off delete my conversations?
No, and that is the most common misunderstanding here. The two are separate controls doing separate jobs: one governs what future material feeds the model, the other removes what is stored on your account. If you want both, use both.
Disclosure. DearHeart AI may earn a commission if you sign up through a Visit link on this page, at no cost to you. It does not change what we write. How we earn.