AI companion privacy: what services store and how to delete it

People tell AI companions things they would not tell a search engine. The conversations are stored on someone else's servers, linked to an email address and often a payment card. This guide compares what several providers say in their own privacy policies, so you can ask the same questions of any service.

It is built from official documents (sources at the end), checked on October 23, 2026. We have not audited anyone's systems, and a published policy is a statement of intent, not proof of practice.

Across the policies we read, the data falls into six groups:

Data type	Example from an official policy
Account	Email, username, date of birth (PolyBuzz)
Conversations	Inputs and outputs of chats with AI companions are defined as user content (Candy AI)
Uploads	Photos, voice input (PolyBuzz, Emochi)
Device and logs	Device IDs, IP address, advertising identifiers, browsing logs (PolyBuzz)
Location and contacts	Precise geolocation and contacts are listed as data PolyBuzz may collect
Payments	Name, email, card brand, last four digits, transaction details shared with payment processors (Candy AI)

Some services add age verification. PolyBuzz says UK verification runs through third-party providers and that it keeps only the result, not your ID document.

Often, yes, and the wording matters:

• Candy AI lists "train and develop our AI models and moderation technologies" among its purposes, and says preparing training data "may include human review of de-identified and/or anonymized interactions" (privacy notice, revised October 19, 2026).
• Emochi says uploaded information may be used for "providing services to you and training models to improve our services" (privacy policy, updated June 17, 2026).
• Nomi refers to "training or communications archives" and says that after account deletion, information in training archives "would no longer be attributable to you" (privacy policy, updated April 27, 2026).
• PolyBuzz asks consent to use uploaded photos to "train our AI algorithm" and shares data with third parties, including "the Chat-GPT platform", to deliver functionality (privacy policy).

Two takeaways. "Anonymized" or "de-identified" training data is still built from your conversations. And when a service uses an outside model provider, your messages pass through that provider too. Mozilla's 2024 review of 11 romantic chatbots found that most policies gave little information about how conversation content was used for training (Mozilla Foundation).

"Private" usually means other users cannot see it. It rarely means the operator cannot. PolyBuzz's FAQ calls chats "completely private", while its privacy policy treats input text as collected data shared with service providers. Candy AI's policy allows human review of de-identified interactions. None of the policies we read promise end-to-end encryption of chats.

Assume that staff, contractors, model providers and, in a breach, strangers could read what you type.

Deleting an account and deleting data are not always the same thing:

• Candy AI keeps account data for three years after last activity by default, or less, and deletes it "without undue delay" if you request deletion. Financial records are kept ten years for tax law; marketing data up to two years.
• PolyBuzz deletes accounts via Setting → Account → Delete, but backups are cleared only when the backup system is updated, and anonymized data and legally required data are kept. Details are in Is PolyBuzz safe?
• Nomi says deleting your account deletes all personal information except what sits in training or communications archives or is needed for legal proceedings.
• CrushOn AI's official FAQ says an active subscription can prevent account deletion.

In practice: cancel the subscription, then delete the account, then email the privacy contact asking for confirmation of deletion. Steps for cancelling are in the billing guide.

In September 2024 the companion site Muah.AI suffered a data breach. Have I Been Pwned lists 1.9 million email addresses exposed alongside the prompts users had written, a breach first reported by the tech news site 404 Media in October 2024. Have I Been Pwned classifies it as a sensitive breach, so it is not publicly searchable; only verified owners of an email address can see whether they were included.

The lesson is not about one company. Prompts were stored in a form that could be linked to email addresses, and an email address is often enough to identify a person. Any service that keeps chat logs carries the same risk. We do not recommend Muah.AI and do not link to it.

Weak security is common in this category: Mozilla's 2024 review found that 10 of 11 romantic chatbots failed its minimum security standards, such as requiring strong passwords.

1. Find the operator. The privacy policy should name a company and a country. No name, no signup.
2. Search the policy for "train". Note whether there is an opt-out and whether it covers chats or only uploads.
3. Find the deletion path and any condition, such as cancelling the subscription first.
4. Check retention periods. "As long as necessary" is not a period.
5. Use a separate email address with no real name in it, and a unique password.
6. Do not upload photos of yourself or anyone else, and never share details about real people.
7. Refuse permissions the chat does not need: contacts, precise location, microphone.
8. Choose the payment route deliberately: app store purchases leave a record in your Apple or Google account; direct card payments may use a neutral descriptor.
9. Search the service name with "breach" before you pay.
10. Write down where you subscribed, so you can cancel and delete later.

Every brand review on this site has a privacy section with the operator's own wording and a link to the source.

• Candy AI Privacy Notice (revised October 19, 2026)
• Emochi Privacy Policy (updated June 17, 2026)
• Nomi Privacy Policy (updated April 27, 2026)
• PolyBuzz Privacy Policy (updated June 1, 2026)
• PolyBuzz FAQ
• CrushOn AI official FAQ
• Have I Been Pwned: Muah.AI breach entry
• Mozilla Foundation: Privacy Not Included, romantic AI chatbots (February 2024)

All sources were checked on October 23, 2026. Policies change; the date of each document is given where the provider publishes one.