Most AI chat apps collect four things: the content of your conversations, whatever account details you hand over, device and advertising identifiers, and usage analytics. For cloud-based apps the messages themselves are the significant item, because they’re sent to a server, stored under your account, and may be reviewed by humans or used to improve the service.
You can check almost all of this before installing anything. The store listings are required to declare it, and there’s a two-minute test that reveals more than any policy document.
The categories, and which ones matter #
| Data type | What it usually includes | How exposed it is |
|---|---|---|
| Message content | Everything you type and the replies | The sensitive one for companion apps |
| Account identifiers | Email, phone number, social login | Ties conversations to a real identity |
| Device identifiers | Advertising ID, device model, OS version | Used for ads and cross-app tracking |
| Usage analytics | Screens visited, session length, features used | Low sensitivity, near-universal |
| Purchase data | Subscription status, receipts | Handled by Apple or Google, not the app |
| Photos and contacts | Only if you grant permission | Check why an app wants them |
For a companion app the ranking is not subtle. Chat content plus an account identifier is the combination that turns a private conversation into a record attached to your name. Analytics and crash logs are, by comparison, boring.
How to read the store privacy labels #
Both app stores publish developer-declared disclosures, and they’re the fastest way to compare apps.
On Google Play, the Data safety section lists what the app collects and what it shares, and developers have to declare each data type as required or optional, say whether data is encrypted in transit, and say whether users can request deletion. Google’s own Data safety documentation lists the categories developers must declare, including messages, photos, device IDs, app activity and personal info. Note that “collect” there has a specific meaning: transmitting data off the device. Anything that stays on your phone isn’t collected, and a well-built on-device app can honestly declare very little.
On the App Store, the App Privacy card does the same job with three groupings: data used to track you, data linked to you, and data not linked to you. “Used to track you” is the one to read first, because it means identifiers are shared with third parties for advertising.
Two cautions. These labels are self-declared by developers rather than audited, so they’re a starting point, not proof. And a short label can mean either a genuinely private app or a careless disclosure, which is why the airplane-mode test below is useful.
What the privacy policy tells you that the label doesn’t #
The label covers what’s collected. The policy should cover what happens next, and these are the clauses worth finding:
- Retention. How long messages are kept, and whether deleting a chat deletes the server copy.
- Training. Whether your conversations are used to train or fine-tune models, and whether you can opt out. This is the clause that matters most for anything you’d be embarrassed to see reproduced.
- Human review. Many services have staff or contractors read a sample of conversations for safety and quality.
- Sharing. Which third parties receive what. Ad networks and analytics providers are normal; data brokers are not.
- Deletion. Whether there’s an actual delete-my-data route, and whether it covers backups.
- Jurisdiction. Where the servers are and which law applies.
If a policy doesn’t answer the training question, assume the answer is yes.
The airplane-mode test #
The quickest way to tell whether an app is really processing your chats locally:
- Install it and complete setup, including any model download it asks for.
- Turn on airplane mode.
- Send a message.
If you get a real, contextual reply, the model is running on your device. If it spins or errors, your messages go to a server. This test cuts through marketing language in about thirty seconds, and it’s worth running on anything that advertises itself as private.
Be aware of a middle case: some apps run small tasks locally and send harder ones to the cloud. A reply in airplane mode proves local inference is happening, not that nothing is ever sent, so pair the test with the store label.
What “on-device” actually changes #
When the model runs on your phone, the chat content never leaves it, so there’s no server copy to breach, subpoena, review, or train on. That’s a structural difference rather than a policy promise, and it’s the reason the category exists.
Xin is built this way: the language model is a file on the phone, and the conversations, the characters and the saved memories sit in a local database beside it. There’s no sign-up, no login and no email, so nothing ties a conversation to your identity in the first place, and “Erase all data” in the Me tab removes the lot from the device. Its network traffic is limited to three things it’s upfront about: the one-time model download, ads on the free tier, and purchase receipts through the App Store or Google Play.
That last point is the honest caveat for any free app, including this one. Ad-supported tiers involve an advertising identifier and a consent prompt, which is data collection even when your messages aren’t part of it. If you want zero ad identifiers, a paid tier with ads removed is the only version of any app that gets you there. The broader trade-offs are in private AI chat apps: what “private” really means.
A short checklist before you install #
- Read the Data safety or App Privacy card, focusing on whether messages and identifiers are collected.
- Search the privacy policy for “train,” “retention” and “delete.”
- Check whether an account is required. No account means much less to link to you, as covered in AI chat apps without an account.
- Run the airplane-mode test after setup.
- Decide what you’re willing to type. Even with a good app, there’s no reason to give a chatbot your full name, address, employer or financial details.
Frequently asked questions #
Do AI chat apps use my conversations to train their models? #
Many cloud services do, though policies vary and some offer an opt-out or a setting that excludes chats from training. Check the policy specifically, because this is rarely stated on the store label. Apps that run the model on your device have nothing to train on, since your messages never reach them.
Can employees read my AI chats? #
At cloud services, some conversations are reviewed by people for safety and quality, and most policies say so. Access is typically limited and logged, but it exists. On-device apps have no such access because there’s no server copy.
Does deleting a chat delete it from the company’s servers? #
Not always, and not immediately. Deleting a conversation often removes it from your view while copies persist in backups or safety logs for a set retention period. Saved memories are frequently stored separately from the transcript and survive a chat deletion entirely, which is why making an AI chatbot forget something takes more than clearing the history.
Are AI companion apps worse for privacy than other apps? #
The data isn’t unusual in kind, but it’s unusually sensitive in content. People tell companion apps things they wouldn’t put in a text message. That raises the stakes on retention, training and breaches, which is why the on-device option is worth considering for this category specifically. Our full risk rundown is in are AI companion apps safe.
What does “we don’t sell your data” actually mean? #
Less than it sounds. It usually means no direct sale of personal information, while still allowing sharing with advertising and analytics partners, which many privacy laws treat as a sale in some circumstances. Read the sharing section rather than the headline.