You have two sign-up screens open and one of the apps is free. Somewhere behind that button is a sentence about whether the conversations you are about to have will be used to improve the model, and it is the line almost nobody reads. AI girlfriend app training data is not a scandal and it is usually not hidden. It is a term of trade, like the message allowance and the renewal date, with a price attached — sometimes to your wallet, sometimes to what you are willing to type.
Two different things get called learning
Two processes on these apps are both described as the app learning, and they have almost nothing in common.
The first is personalisation. The app keeps notes on your character and the things you have told it, and feeds them back into later conversations so the thread holds together. That stays inside your account, and on most apps it is the biggest thing the paid tier is selling. Our guide to how memory works and what the paid tier adds covers that side.
The second is model training: your conversation text, normally stripped of obvious identifiers, added to the pool of material used to improve the shared model every user talks to. That one leaves your account by definition. An app can do the first without the second, and plenty do. A reader who has decided they are comfortable with memory has not thereby decided anything about training, and the paywall screen is happy to let the two blur together.
Where AI girlfriend app training data shows up before you sign up
All of this is checkable without an account, one of the few privacy questions you can settle while still comparing.
The store listing is the quickest look and the most limited. Each App Store product page carries a privacy section listing the categories an app says it collects, and whether they are linked to you or used to track you; Apple's note on privacy information on the App Store explains those two headings. Notice what is missing: no category covers whether your messages feed a training run. The labels answer the advertising question, not this one.
So the policy is where you have to go. Open it in a browser and search the text for train, improve, machine learning, aggregated and de-identified. What you want is a sentence saying plainly whether conversation content is used to develop models, and whether you can decline. What should slow you down is a policy that details device identifiers and analytics precisely, then covers the chat log with a phrase like "to improve our services" — doing a great deal of quiet work in an otherwise specific document.
Then the help centre, usually indexed and readable from outside. If an opt-out exists, someone has written a page about where the toggle lives, and it often answers something the policy does not: whether the control is on every plan or only a paid one.
What an opt-out actually changes
Finding the toggle is not the end of it, because these controls are narrower than they read.
They are generally forward-looking. Turning training off typically stops future conversations being used; it does not unpick material already in a dataset, and no app can promise to pull your sentences out of a model already trained on them. Which is why the toggle is worth finding on day one rather than in month three.
They also tend not to cover safety and moderation. Reviewing flagged conversations for abuse, age-policy breaches and illegal content is a separate lane from model development, often required of the company, and rarely something a user setting switches off. The honest assumption is that a conversation could be read by a person under some circumstance, whatever the training setting says.
Regulators are now asking this directly rather than inferring it. The US Federal Trade Commission's inquiry into AI chatbots acting as companions, opened in September 2025, required companion-product makers to report on how they monetise engagement and how they use and share personal information obtained through users' conversations. Nothing in it tells you what any one app does. It does mean today's disclosure is written under more scrutiny than one from two years ago — read the current version, not a summary of an old one.
A training opt-out is a promise about tomorrow's conversations. Nothing you have already typed is covered by it.
The part of this that is priced
Here is where it meets the comparison you are making. On free and cheap tiers your conversations carry more weight in the business model, because the subscription is not carrying it. That is not sinister, and a free tier is not a trap for being free. It is the reason to read the terms most carefully where the price is lowest.
It also explains a pattern worth pricing honestly: on some apps the privacy controls, training opt-out included, sit on a paid plan. Then the control is not free, and belongs in the same column as the message allowance. Take the monthly difference, multiply by the months you realistically expect to use the app, and weigh it against the alternative of writing less revealing messages on the free tier. Our piece on what paying for privacy settings actually buys works through that arithmetic.
One thing it does not cost is quality. Declining training leaves your own memory and personalisation untouched, so there is no performance reason to keep it on — only the time to find the setting, and on a paywalled toggle, money.
Which reader should weigh this heavily
AI girlfriend app training data matters much more to some readers than others. Three shapes, and most people know which they are.
- You write things you would not want read. Details about your life, your work, people you know. Treat a documented opt-out as a requirement rather than a bonus, pay the difference if it is behind a plan, and still write as though a person could see it.
- You keep it fictional and light. Invented scenarios, no real names, nothing about your job. Training is close to academic for you; tracking — advertising identifiers, data shared with brokers — is the one with consequences. Read the store label rather than the policy.
- You are going to pay anyway. Check whether the privacy controls arrive with the tier you were going to buy or only the one above it. It is the detail most likely to decide between two apps you otherwise rated equally.
The ten minutes before you sign up
A short sequence, in the order that saves most time:
- Read the store listing's privacy section for the tracking answer.
- Search the policy for the five words above and read the sentences around each hit.
- Search the help centre for the opt-out and note which plan it lives on.
- Sign up with an email address you use for nothing else, which limits what this can be joined to later.
- Set the toggle before the first conversation, not after the first month.
If that is more work than you want to do on two apps at once, do it on one. A few unhurried weeks on the free tier of the app we currently recommend, settings checked on day one, tells you more than an afternoon of comparing documents.
Frequently asked questions
Do AI girlfriend apps use my chats to train their models?
Some do and some do not, and the only reliable answer for a given app is in its current privacy policy. Search the document for "train", "improve" and "machine learning" rather than trusting a summary, including this one — policies get updated without announcement.
Does turning off training delete what the app already has?
No. A training opt-out normally applies to future conversations only, and it is a separate control from deleting your data. If you want the history gone too, that is a deletion request — see our guide to deleting an account rather than just the plan.
Is memory the same thing as AI girlfriend app training data?
No, and the difference is the useful part. Memory is personalisation held against your account, so your own conversations stay consistent. Training is conversation content used to develop the shared model all users talk to, which means it leaves your account.
Disclosure. Noxlox 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.