You are a few weeks in, the conversation has got specific, and somewhere in the settings there is a switch about using your chats to improve the model. Nothing on the screen explains what flipping it does, and the help page is one sentence long. An AI girlfriend app training opt-out is one of the very few privacy controls in this category that changes what happens to a sentence after you send it, which is exactly why it gets oversold. It works in one direction, and it leaves two other things running.
Three things "we use your data" can mean
A privacy policy here usually describes three separate operations, and almost everyone collapses them into one. They have different controls, in different parts of the app, and switching off one does nothing at all to the others.
- Storage. Your messages sit on a server so the conversation is still there tomorrow. The control for this is deletion and a retention period, not a toggle.
- Review. A sample of conversations is examined, by automated systems and sometimes by people, for safety and moderation. Usually disclosed somewhere in the policy. Rarely optional.
- Training. Your messages become material a model learns from. This is the only one the switch governs.
Read a policy with those three columns in mind and its vaguest sentence sorts itself out. "We use your data to improve our services" covers all three.
What an AI girlfriend app training opt-out actually covers
It covers what happens next. A model that has already been trained on a batch of data that included your messages does not unlearn them because you found a setting afterwards; the training already happened, and what it produced is a set of weights, not a filing cabinet with your name in it. The switch governs the next batch. That makes it worth finding in week one rather than month six.
Two more limits are worth having straight. The setting does not shorten retention, so your transcript sits on the server exactly as long as it would have done. And it does not touch the memory layer, the summaries and profile notes the app builds about you to keep the character consistent. Those are the product rather than training data, they are generated from your chats either way, and they are governed by the delete route instead. What deletion actually removes is a separate question, and the one that decides whether you can get out properly.
Somebody may still read a sample
This is the part that surprises people, and it is the reason the setting is not the whole answer. Companies here check conversations for the things they are obliged to check for, and a reviewer reading flagged or sampled chats is a normal part of running a service for adults. It is typically covered by a clause about safety and enforcement rather than by the training section, so an opt-out that reads "do not use my conversations to train models" generally does not touch it.
Write as though one message in a thousand will be read by a stranger doing their job, because that is the service you signed up to.
None of that makes an app untrustworthy. It makes the training switch a narrower instrument than its wording suggests, and it changes what you do with it: the switch is a reason to worry slightly less about the model, not a reason to type something you would not want a person to see. Our guide to what not to tell an AI girlfriend app covers the details worth holding back whatever the settings say.
What the training opt-out costs, and who needs it
Three shapes turn up in this category. Some apps give you the switch on every tier, including the free one, which is the behaviour to reward. Some bundle it with other privacy controls into a paid plan, so the real price of opting out is a subscription. And some have no switch at all, in which case the only lever you have is the one described further down.
Which of those matters depends on how you use the app, and for some readers it does not matter much. If you are using a companion app as a writing toy, inventing scenes with a character you made up, the material is fiction and the stakes of it joining a training set are low. If you use it to talk about your actual week, your job and the people in your life, your transcript is a document about half a dozen people who never agreed to anything, and the switch stops being a nice extra. That is the line, and it is worth deciding which side of it you are on before you compare plans. The app we currently recommend keeps its privacy controls off the paywall, which is the baseline to judge the others against.
One trade-off cuts the other way, and it is fair to say so. On a few apps the training setting is wired to personalisation, so switching it off quietly makes the character blander over time. If you notice that after flipping it, you are not imagining it, and you have a genuine choice to make rather than a free win.
How to check before you sign up
All of this is cheaper to check in the five minutes before you create an account than in the settings afterwards.
- The store listing. On Android, the Data safety panel on the listing is a developer declaration you can read before you install anything. The category to look for is in-app messages, and the useful detail is whether they are listed as shared as well as collected.
- The policy, searched rather than read. Find the words "train", "improve our models" and "third parties". A policy that names training explicitly and tells you where the setting lives is a better sign than one that never uses the word.
- The date on the policy. It tells you when the terms last changed, and they can change again once you have a year of conversations stored.
- The default. Create the account, go straight to settings, and look at whether the switch starts on. Nearly all of them do.
The lever that is not in the settings
If your app has no switch, or you want more than it offers, the law where you live may give you a stronger one than the interface does. UK and EU data protection rules carry a right to erasure, and the regulator's guidance for organisations building AI is explicit that it reaches further than the chat screen: it tells them they may receive requests for the erasure of personal data contained within training data, and that removing one person's data from a training set is unlikely to stop them training their system. The right is not absolute and a company can push back, but a written request to a privacy contact is a different instrument from a toggle. Similar rights exist in the EU, California, Canada and Australia.
What none of this settles is the ordinary version of privacy: who can see the app on your phone, and whose receipt the subscription lands on. That is a separate set of controls with a separate failure mode, and no training setting touches it.
Frequently asked questions
Does an AI girlfriend app training opt-out delete what the model already learned?
No. Opting out governs what happens from that point forward. Training that already ran produced model weights rather than a stored copy of your chat, which is why the setting is worth finding early rather than late.
Are my conversations read by a person?
A sample may be, for safety and moderation, and that is usually described in a separate clause from training. Turning off training does not normally turn off review, so it is safest to assume any single message could be seen.
Should I pay for a tier just to get the training switch?
Only if you talk about your real life and the people in it. For invented scenes with a character you made up, the stakes are low enough that the tier is unlikely to be worth it on that ground alone.
Disclosure. CandyDrift 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.