Do AI girlfriend apps read your chats? Moderation, training and staff access

Four ways your messages are read

1. The AI itself. To reply, the model has to read your message. Often that model belongs to another company. Candy AI's privacy notice says third-party LLM providers may receive the content of your messages; Anima names OpenAI and Groq among the AI providers that receive your conversations; Replika's policy says conversation data is sent to third-party language model providers after de-identification.
2. Automated moderation. Apps scan messages for banned content. Candy AI says chats are scanned by its own moderation system, and Lovescape says all messages and prompts pass through automated moderation.
3. Human review. Flagged chats can reach staff. Candy AI mentions possible manual review of flagged content, Lovescape says human review follows three or more violations, and Roger and Swipey AI say flagged chats may be reviewed by staff.
4. Training. Chats can be used to improve the AI, sometimes after names are removed.

This is not unique to companion apps. In September 2026, 404 Media reported that OpenAI was hiring hundreds of contractors to read real ChatGPT prompts, with usernames removed.

What each app says about training and access

From each app's own privacy policy, as recorded in our data (September 2026):

Where a privacy policy says nothing about training, that is not a promise it never happens. Our Nomi AI review and OurDream review cover their policies in more detail.

End-to-end encryption claims

End-to-end encryption means only the sender and receiver can read a message. In an AI chat, the receiver is the company's software, so the company can always read what you send it. The claims in our data:

• OurDream's homepage says no one can read your chats, "not even us", while its own privacy summary says third-party model providers process each message. Its policy describes AES encryption at rest and TLS 1.3 in transit.
• Secrets AI's homepage claims end-to-end encryption; its privacy policy describes AES-256 encryption of stored data and TLS 1.2 or later in transit.
• Nomi AI's help wiki says plainly that end-to-end encryption is not practical for an AI chatbot, so chats are not end-to-end encrypted.

Encryption at rest and in transit is still worth having: it protects your chats from outsiders. It just does not stop the company reading them.

Why it matters: leaks and licenses

Stored chats can leak. In 2025, two companion apps left 43 million messages from over 400,000 users on an unprotected server, and a 2026 security audit of 17 AI companion apps found that 10 had routes to stored conversations. Our guide to AI girlfriend data breaches lists the known cases.

Terms of service matter too. Common Sense Media's 2025 report found that companion platforms' terms grant extensive, often perpetual rights to what users share, quoting Character.AI's terms, which at the time granted a perpetual, irrevocable license to use content for any Character.AI-related purpose.

How to keep your chats more private

• Opt out of training where the app allows it. Character.AI, for example, lets users in the UK and EEA switch off its model-improvement setting.
• Share nothing identifying. No full name, address, employer or photos of real people.
• Use a separate email and a strong, unique password.
• Delete chats you would not want read, where the app lets you delete individual chats.
• Delete your account when you leave. Our guide on how to delete an AI girlfriend account covers each app.

For the wider picture, see whether AI girlfriend apps are anonymous.