AI companion privacy: chat data, retention and deletion

Start with the information you intend to share, then check how the service handles that information. AI companion apps can collect personal information across chats, memory, uploads and account activity. These privacy risks need separate checks: a private character, encrypted connection and cancelled subscription answer different questions. This guide gives you a data map, examples from published policies and a record you can use before enabling chat, memory, uploads or voice.

Our scope is the provider’s public explanation and the controls a reader can inspect. We have not audited the services’ security or tested their deletion systems. The policy examples below were checked on September 17, 2026; they are examples of how to read a disclosure, not a ranking of the safest AI companions.

AI companion data collection: map what the chatbot receives

Begin with one intended activity, such as continuing a fictional conversation. List the information needed for that activity before adding optional features. A text chat does not necessarily require your real workplace, a photograph of another person or access to your contacts. Combining ordinary details across many conversations can make an account more identifying than its nickname suggests.

Mark a category as unused when it is outside your evaluation. Mark a missing answer as unclear. Neither label is a finding that a platform is secure or unsafe; it keeps the comparison tied to the activity you are actually considering.

AI companion privacy risks: six policy questions

1. Collection: does the explanation cover messages, uploads, generated material, account identifiers and usage logs?
2. Purpose: distinguish running the service from moderation, support, analytics, advertising and model improvement.
3. Access: look for staff, contractors, infrastructure providers and situations such as a reported support problem.
4. Retention: identify what starts the clock: account creation, inactivity, a deletion request or the end of a subscription. Read the exceptions too.
5. Training: look for an explicit description of whether content is used to train or improve models and any control over that use.
6. Removal: separate the visible interface action from the provider’s explanation of retained copies, memories and billing records.

Visibility and encryption are different parts of the data path

Private chat can mean that other users cannot browse the conversation. It does not establish that the provider never stores or processes it. HTTPS protects the connection in transit; the browser’s lock icon does not tell you who can access stored records. Claims about encryption at rest or end-to-end encryption need their own explanation of keys, endpoints and exceptions.

Read the marketing page and the policy together. When they appear to describe different limits, save both URLs and ask about the specific feature. Do not resolve the ambiguity by choosing the more reassuring sentence. Likewise, a friendly AI character’s promise of secrecy cannot establish a contractual or technical privacy commitment.

Three policy examples that change the next question

These examples show why one broad privacy label is insufficient. They describe published wording, not independently observed data handling. The services may revise their policies, and a policy can leave details of individual features unresolved.

The useful outcome is a narrower question. For example, knowing that a policy mentions retention does not tell you that every data category follows the same duration. Knowing that support may review a reported conversation does not establish unrestricted access in every situation. Preserve the condition attached to the statement.

Data protection controls: account access and public visibility

Before creating substantial content, locate the account settings, sign-in methods, available session controls and official support route. Use account-protection options the service actually provides. If you do not find a particular setting, record that observation without claiming the feature does not exist anywhere.

Review character visibility separately. A fictional name can still sit above a description containing real events or identifying details. Inspect the greeting, biography, images and sharing settings before publishing a character. Private, unlisted and public labels should be interpreted using that platform’s documentation rather than assumptions from another app.

Check a feature’s input requirements before enabling it. A voice recording, reference image and typed description have different contents and may follow different processing paths. Our OurDream review explains its documented reference-image limitations; that workflow should not be assumed to apply to another provider.

Separate cancellation, deletion, unpublishing and export

For a small export check, use a disposable fictional conversation, open the resulting file and compare it with what you expected. A successful download proves that a file was produced, not that it includes every category. Protect the exported copy: moving data out of an account creates a separate storage responsibility.

Before an irreversible action, retain necessary receipts and original work. Read the exact confirmation. A web subscription and an app-store subscription can have different controls, even when the product name is the same. Save cancellation confirmation and check the displayed renewal state through the channel you used to pay.

Evaluate the experience with a fictional profile

Use an original adult character such as Rowan, a 34-year-old radio producer in an invented town. Give Rowan a harmless preference and a fictional appointment. Continue an ordinary conversation, correct the appointment and later ask about the new date. The exercise evaluates handling of supplied context without making your private life the test material.

Keep the same invented facts when comparing two services and record the actual replies. Our AI companion memory guide provides a fuller protocol. Correct recall is a conversation observation; it does not reveal where information was stored or whether it would be removed after deletion.

Start with text if that answers your question. Add uploads, voice or public sharing only when those features matter to you and you have checked their handling. Mark an unused feature as untested. You can reach a useful buying decision without using every optional mode.

Keep a dated record and ask one precise question

A short record makes policy comparisons easier to revisit. Copy the following fields into a local note, with one entry for each provider and feature. Avoid pasting sensitive conversation excerpts into the record when a description is enough.

• Feature and data: text chat; fictional character facts only.
• Source: policy URL, policy date if displayed and the date you checked it.
• Documented: the relevant statement, including its condition or exception.
• Observed: the control you saw and what you checked, with no claim about hidden backend behavior.
• Unclear: the question the source does not answer.
• Decision: what you will use, postpone or leave out until the question is resolved.

For example, a focused support question is: “If I delete one conversation, are the extracted memories removed too, or must I delete them separately? Please point me to the applicable instructions.” This is more answerable than asking whether the whole service is private. Send it through the provider’s verified support route without including a password or unnecessary personal material.

Recheck the record when you enable a different feature, change visibility or receive a material policy update. An old screenshot is evidence of an old interface. Keep unresolved points explicit rather than filling them with assumptions.

Understand what a referral link adds

A commercial visit link on NSFW Guides may pass through an affiliate network before reaching a provider. That referral is separate from the later AI conversation. Our affiliate disclosure explains the commercial relationship; the network and destination service have their own data-handling terms.

Check the destination domain and offer before entering information. A link to a policy in this guide is a source citation. A labeled visit button is a commercial referral. Neither the referral nor our editorial analysis certifies the destination’s security.

AI companion privacy questions answered

What should you ask about third-party processing?

Look for the provider’s explanation of the services used for hosting, model processing, analytics and billing. Identify which data category each disclosure covers and whether optional features change the path. A general claim that a platform uses artificial intelligence does not tell you which organizations receive a chat, upload or account identifier.