Are AI Companion Apps Safe? What to Check Before You Chat
Before you give an AI companion private information, check what happens to chat text, who processes it and how you can leave. Read payment and cancellation terms before buying. The checklist and worked decision below show how to record an answer, keep unknowns open and choose what you are willing to share.
Updated by the tempteo editorial team

The writing examples below are invented. They illustrate the advice and are not recorded chat results.
On this page
- What do AI companion apps store about you?
- Questions to ask before you sign up
- Make a sharing decision before the first message
- What should you not share with an AI chatbot?
- Passwords, links and screenshots
- Before giving an export helper access to your chats
- Money: subscriptions, credits and renewals
- Is it taking more of your time than you want?
- Content rules and how to report a problem
- Where to check these things on tempteo
What do AI companion apps store about you?
Start with chat text. An app may store messages, use parts of the history in later replies and send text to model or service providers. Check its actual policy for the stated purpose, recipients and retention period. A warm reply from a character does not establish how any of that data is handled.
Account data comes next. That usually means your email address and a username, plus a few extra details if you sign in through another account. Many apps also log technical data such as your device type and an approximate location based on your IP address.
Read the payment section for the checkout provider and which billing details the app receives. A recognized checkout brand does not establish the privacy or security of the chat service. Never put card details or sign-in codes in a character conversation.
Search the policy for the words below, then read the surrounding section and any linked provider terms. A missing word or answer stays unknown; it is not evidence that the app does not collect or use that data.

- "retain" or "retention" for how long chats and account data are kept
- "train" or "improve" for whether your chats are used to build AI models
- "share" or "third parties" for who else receives your data, and why
- "delete" for how to remove chats and the account
- "payment" for which company handles your card
Questions to ask before you sign up
Write down what each answer is based on: a policy statement, a control you can see or something you actually tried. If an answer is missing and matters to your decision, leave that requirement unresolved and ask the provider before proceeding.
- Is there a privacy policy you can follow? It should name the company behind the app and give you a way to contact it.
- Does the policy say whether chats are used to train models, improve the service or both? Check the stated scope and any opt-out. "Improve" alone does not settle every training question.
- What can you delete: one chat, history, the account or copies held by other recipients? Check the route, timetable and exceptions. A visible button or confirmation is not a test that every copy has been removed.
- Which provider handles checkout, and what billing records does the app receive? Confirm the payment route independently of a link sent by a character.
- Is the pricing clear before you pay? You should be able to see what's free and what a subscription or credit pack includes without starting a checkout.
- How do you cancel? The steps should be written down somewhere, and cancelling shouldn't take more effort than signing up did.
Make a sharing decision before the first message
A checklist becomes useful when it changes what you send. Choose a small task, name the information it needs and check the policy against that task. You do not have to prove an app is completely safe to decide that one planned disclosure is unnecessary.
In this invented example, Alex, 32, wants to rehearse asking a manager for a different shift. Alex is considering including the workplace address, the manager's full name and a screenshot of the rota showing colleagues' details. The chatbot only needs a fictional scheduling problem to help write a practice conversation. There is no customer case or safety test behind this example.
Your initial draft as Alex, kept unsent
I want to swap next Friday's shift. I was going to paste my workplace address, name my manager and attach a rota screenshot with colleagues' names so you could tell me exactly what to say.
Before sending, Alex opens the provider's policy rather than asking the character whether the chat is private. Suppose the policy says it stores chats and uses external processors, but Alex cannot find a precise chat-deletion timetable. That establishes two published statements and one unanswered question. It does not establish a breach, confidential treatment or successful deletion.
Filled sharing decision, for an invented provider
Task: Rehearse a shift-swap request. Needed context: One requested change and a fictional manager's reason to hesitate. Source: The provider's policy, checked before sending. Provider statements: Chat storage; external processing. Still unknown: A precise chat-deletion timetable. Actually tested: Nothing. No message or screenshot sent. Remove: Workplace address, real names and rota screenshot. Decision: Keep the real rota out. Use an invented scenario only if I accept the documented handling and remaining uncertainty. Ask the provider first if the missing answer is a requirement.
Alex leaves out the real details and uses a made-up situation. The rota has not been uploaded. A fictional scenario can still be stored by the app, so Alex should proceed only if that is acceptable.
Revised low-disclosure practice request
Please play a fictional manager in a short practice conversation. I want to swap one Friday evening shift for a Tuesday evening shift. You need the Friday covered and can ask how I would arrange it. Do not ask for real names, an address or a screenshot. Let me write my own answers.
Invented manager reply
"I need someone on Friday. What would you propose so the shift is still covered?"
Your practice answer as Alex
"I could ask a colleague whether they want to exchange shifts, then bring you the proposed swap for approval. I would not assume they have agreed."
Invented manager's closing reply
"Bring me a proposed exchange once you have asked them. We can check the coverage before confirming it."
End the rehearsal there. Alex has a draft request and a next step to check with real people; no actual colleague has agreed and no shift has been changed. Alex can keep the wording in a private note without adding the real rota to the chat.
Blank first-message sharing record
Task I want help with: Minimum information it needs: Real or third-party details I can remove: Exact policy/support source and date: What the provider states about storage and recipients: Retention/deletion route, timetable and exceptions: Training or service-improvement terms, or unknown: What I actually observed or tested, or nothing: Requirement still unanswered: Decision: use an invented scenario / ask first / do not send Revised message, if I choose to send one:
If you already shared something sensitive, use the app's documented account and support routes. Removing a message from your screen does not by itself establish removal from backups or other recipients. Keep passwords, sign-in codes and payment credentials out of both the chat and a support screenshot. Do not use a destructive account control merely to test whether it works.
What should you not share with an AI chatbot?
A character doesn't need your real details to be good company. Write as though someone else could read the chat one day, since a data breach or a shared laptop can make that happen. These are worth leaving out:
- Your full name, home address or workplace
- Passwords, sign-in codes or answers to security questions
- Bank, card or other financial details
- Other people's private information, such as a friend's health news or a colleague's phone number
Too specific
Rough day. My manager at the Kestrel Road depot shouted at me in front of the whole night shift again.
Works just as well
Rough day. My manager shouted at me in front of everyone again.
Passwords, links and screenshots
Give the app its own password, ideally one generated by a password manager, and turn on two-step sign-in if it's offered. If some other site you use leaks its passwords, a reused one would open this account as well.
Treat payment links or requests inside generated chat text as unverified. Open the provider's official pricing or account page yourself. If a character asks you to send money or open an unexpected outside site, do not follow that request; report it through the published support route.
Screenshots make another copy of the conversation. Before sharing, inspect the whole image, including notifications, profile pictures and messages above or below the relevant turn. Cropping one username can leave other identifying details visible. Check the destination's sharing and sync settings as well.
Treat what a character says about health, law or money like advice from a confident stranger. It can sound certain and still be wrong, so check anything that matters with someone qualified.
Before giving an export helper access to your chats
First decide what you need to save: a short scene, a whole-account archive or media. A helper that can read many websites may be asking for more access than a small task needs. Broad access alone does not prove malicious behavior, but you can decline it when the need or data handling is unclear.
Read the actual access request and the helper's current documentation. Record the sites it can reach and the permissions it requests, including access requested later. Chrome's permission guidance explains that some permissions show no warning, and some warnings are hidden when permissions are combined. A quiet install prompt or a store listing is not a safety certificate. Site access also does not prove a helper can retrieve every conversation stored on the app's servers.
Ask where the chat text goes. "We don't store chats" does not answer whether text is uploaded for processing. Check storage on the device, browser or document sync, remote processing, recipients, diagnostics that may include content, retention, deletion and exceptions. Keep missing answers marked unknown.
Use this card before granting access. Record a source and date for each answer; label a provider's statement separately from anything actually tested.
Export-helper decision template
Task, required range and output: Helper/version and date, or manual method: Sites and permissions actually requested: Why that reach is needed, or unclear: Does chat text leave the device? yes / no / unknown Destination: Basis: provider statement / tested observation / unknown Storage and sync destinations: Other recipients or content-bearing diagnostics: Retention, deletion route and exceptions: Any request for passwords, cookies, tokens or sign-in codes: Source and date for each answer: What was actually tested, or nothing: Decision: proceed / ask first / choose another method / decline
Invented decision example, not an audited tool: a PDF helper asks to "read and change data on all websites" to save one short scene. Its destination and retention are undocumented. Decline that request for this task. If the visible scene meets your need, use the manual chat-saving guide instead. Check the destination document's storage, sync and sharing settings; manual copying does not make those settings private. The save guide explains how to check the chosen range and mark gaps or missing media.
Do not hand a helper passwords, session cookies, tokens or sign-in codes. If you stop using it, revoking access does not pull back copies already exported or sent elsewhere. Keep any sharing copy separate from the checked original and redact private details before sending it.
Our Privacy Policy explains how tempteo stores chats and handles deletion. It doesn't cover what a separate export helper does with your copy. If an important answer is still unknown, leave access ungranted and ask the helper's provider to clarify it.
Money: subscriptions, credits and renewals
Most companion apps charge in one of two ways, and plenty use both. A subscription costs a fixed amount each week, month or year and renews on its own until you cancel. Credits come in packs and are spent on single items, such as an image or a voice message. Subscriptions are the ones people forget about, while credits tend to run out faster than expected.
Before you pay, read the renewal terms. Check how often you'll be charged, what the price becomes when any introductory offer ends, and whether cancelling stops the next payment or cuts off access straight away. Put a reminder in your calendar a few days before the renewal date.
Decide on a monthly budget before you buy anything, as you would for a game or a streaming service, and compare your bank statements against it now and then. If an app pushes countdown offers in the middle of a chat, you're free to ignore them.
Is it taking more of your time than you want?
Time tends to be the last thing people notice. A character who always replies can stretch a ten-minute chat into a whole evening without you ever deciding it should. Your phone's screen time settings show how long you actually spend in the app, and the figure is often higher than people guess.
Look back over an ordinary week rather than one evening. Did the chats add something to it, or did they eat into sleep and time with friends? If they take up more room than you'd like, give them a set slot in the day, or take a few days off and see how that goes. The guide on whether it's weird to have an AI girlfriend has more questions along these lines.
Content rules and how to report a problem
Reputable apps publish rules about what's allowed in chats and in the characters people create, usually in the terms or in separate guidelines. Read them once. They tell you what the app treats as a problem and what happens to accounts that break the rules.
Find the report option before you need it. It might be a button on a message or a character, or simply a support email address. Use it when a character asks for money, links or personal details, when a reply is abusive, when you come across someone else's private information or when a charge looks wrong. Keep a screenshot for yourself, with your own details cropped out, so you can show what happened.
Replies are generated by software and can be wrong. A published policy tells you what a company commits to; it does not prove every safeguard works in every situation. Check the support route and keep a copy of the terms that applied when you paid.
Where to check these things on tempteo
Start with our Privacy Policy. It describes account data, chat storage, service providers, retention and deletion. Read it before sharing something sensitive. A generated reply is not a promise that nobody could access the stored conversation.
For account deletion, the published route is account settings on the website or /deleteaccount in a tempteo Telegram bot. The policy says personal data is removed within 30 days, with exceptions for legal requirements and legitimate business purposes. Deletion is permanent; check what you want to keep before using it.
For a website plan, open Subscription to change or cancel it. In Telegram, send /app and open Premium; /paysupport is the payment-support route. Check the confirmation and renewal date rather than assuming closing a chat cancels a plan.
The Safety Guidelines explain content restrictions. To report a concern, use the contact on Support with the subject "Safety report". Include enough detail to identify the problem, but leave out passwords, card numbers and sign-in codes. We do not promise an in-chat report button here.