AI Companion Privacy: What to Check Before Sharing

Correction, September 20, 2026: The earlier guide made unsupported end-to-end encryption claims, generalized how providers train on conversations, and promised complete privacy and deletion from local software. Those claims are withdrawn. This replacement explains what to check and gives dated examples from provider policies. It is not a security audit. Lewdly publishes this guide.

An AI companion can invite personal conversation, but its friendly responses do not explain how the service handles data. Check the provider's policy and your account settings before sharing anything you would not want stored or processed by another organization.

Separate the Privacy Questions

Storage, processing, training and advertising are different uses. A service can retain a conversation to provide chat history without using it to train a general model. It can also send data to a processor to generate a response without selling that conversation to advertisers. Read each claim within its stated scope.

If the policy does not answer a question, treat that point as unresolved. Company size, a paid plan or a reassuring chatbot reply is not evidence of a specific security protection.

Distinguish Transport Encryption From End-to-End Encryption

Encryption during transmission helps protect data while it travels. It does not, by itself, establish that the service operator cannot process or access the conversation. Look for an explicit description of what is encrypted, where processing occurs and who controls access.

For example, Replika's privacy policy, updated May 27, 2026 and checked September 20, describes SSL encryption during transmission and security controls for stored data. It also describes processing conversation data through third-party language-model providers. Those statements do not establish end-to-end encryption that prevents provider access.

That policy distinguishes internal anonymized safety/performance uses from third-party model training restrictions. It also specifies post-termination retention by data category. Account deletion should therefore not be described as immediate removal of every record. Read the current policy for the categories and exceptions relevant to your request.

Read Deletion Exceptions Before Relying on Them

Nomi's privacy policy, effective April 27, 2026 and checked September 20, describes exceptions for information in training or communications archives and information required for legal proceedings or orders. It says information in training archives would no longer be attributable to the user after deletion.

That is a provider's stated policy, not independent verification of its implementation. It also does not establish the earlier article family's claim that all training requires a particular opt-in setting. Keep the exact policy wording and date with any privacy decision you make.

When leaving a service, distinguish cancelling billing, deleting a conversation and requesting account/data deletion. Check each process separately. Save confirmation and contact support about unanswered retention questions without sending unnecessary sensitive information again.

Reduce What You Share

Use a unique password and enable an additional authentication factor if the service offers one. Lock shared devices, review notification previews and check permissions for microphone, camera and photo access.

Provide accurate required account information, but avoid adding unnecessary addresses, financial details, workplace documents or another person's private information to chats. Review images for identifying details before uploading them. A separate display name may reduce obvious identification; it does not make an account anonymous.

These steps reduce particular exposures. They do not guarantee that an account cannot be compromised or that nobody can discover the app is being used.

Verify What Local Processing Actually Means

A local interface is not necessarily a local model. For example, SillyTavern's official documentation describes an interface that connects to language-model backends. Check which backend your installation actually uses before assuming that prompts stay on your device.

For a local workflow, inspect remote connections, extensions, telemetry settings, synchronized folders and backup behavior. Secure the device and accounts that can read the files. Deleting one visible file does not establish that every backup or synchronized copy is gone.

Local inference can reduce transmission when the workflow is configured accordingly. It does not create permanent access, complete anonymity or a guarantee against data loss.

Keep a Dated Privacy Record

Before using a service, note the policy date, selected settings and unresolved questions. Review that record when the provider announces changes or when you enable a new feature, such as voice or image uploads.

Use the official privacy request channel when seeking access or deletion. Do not ask the companion character to confirm backend security or account deletion: a generated response is not an operational receipt. The useful outcome is an informed limit on what you share, supported by current evidence rather than a blanket privacy promise.