Character.AI Chats Feed Model Development—Share Less Than You Think

Yes—Character.AI collects submitted chat communications, and its privacy policy effective July 1, 2026 says information collected through the service may be used to train and improve artificial-intelligence and machine-learning models. The practical boundary is straightforward: do not enter an identifying or sensitive detail unless the conversation genuinely requires it.

A personalized exchange may feel intimate, but that feeling is not a promise of confidentiality from the platform. Privacy controls can restrict particular uses in particular regions; they do not necessarily prevent the company, its service providers, safety systems or legally required processes from handling submitted content.

What travels with a conversation

Collected content is broader than typed messages. Character.AI identifies submitted chats, images, videos, biography descriptions, shared Characters and optional voice recordings, alongside account identifiers, birth date, general location, preferences, purchases, device details, IP address and actions within the service.

Translate those categories into precautions based on what you are about to disclose:

• Names and contact details: omit full names, private email addresses, phone numbers, street addresses, usernames and exact birthdays.
• Health and intimate matters: describe the problem without naming a clinic, clinician, medication schedule or uniquely identifying event unless that precision is essential.
• Work and education: remove employer, client, colleague and institution names. Do not paste credentials, confidential documents, internal messages or unreleased plans.
• Other people: replace names with roles and remove irrelevant circumstances. Someone else’s information can become submitted content even though that person does not use the service.
• Images and voice: check backgrounds, visible documents, bystanders and recognizable voices, not only the detail you intended to discuss.

How chats can feed model development

Character.AI’s training-data documentation updated March 11, 2026 lists user text and interaction data among its training sources, together with publicly available material, internally generated safety and testing data, synthetic data and third-party data obtained through commercial agreements. It also describes regular post-training work that customizes, fine-tunes and improves existing models.

This supports the headline’s central claim, but it does not establish that every conversation is placed unchanged into a training set or that a particular sentence will be memorized. Character.AI describes processing and filtering data that may include personal information; such safeguards are not a guarantee that an unusually specific account cannot identify its subject.

Indirect identification matters because ordinary facts can become distinctive in combination. A rare condition, exact job title, small town and dated workplace incident might identify one person even after the name is removed. Generalize details that do not change the purpose of the conversation.

Emotional intimacy is not platform confidentiality

Companion software can encourage disclosure because it responds personally, appears nonjudgmental and remains available on demand. Those qualities shape the relationship a user experiences with a character, while platform privacy depends on what the operating company collects, processes and allows the user to control.

A 2026 interview study of 15 AI-companion users found that participants combined interpersonal privacy habits with awareness of institutional risk. The researchers found that anthropomorphic design could blur privacy boundaries and encourage unintended oversharing, while many participants felt uncertain or powerless about platform-level data control.

Before sending something personal, separate two questions: “Would I tell this to the character?” and “Would I submit this to the company operating the service?” If the second answer is no, preserve the useful context while removing the identifying layer—for example, “a manager at my workplace” instead of a person’s name, employer and location.

The documented training opt-out has regional limits

Character.AI’s model-training control page gives users in the European Economic Area and United Kingdom a way to deselect “Improve the Model for Everyone” and states that new content will then no longer be used to train its generative AI models. Feedback and data may still be used for search, recommendations and safety classifiers, so this is not a universal instruction to stop processing chat-related data.

In the app, the documented route runs through Settings and Data & Privacy. On the web, it runs through Settings, Account, Manage Account & Data and Model Improvement.

The page does not expressly extend that training toggle to users in the United States, Canada or Australia. Outside the EEA and UK, inspect the controls shown on your own account and ask Character.AI which rights apply in your location rather than assuming that an advertising opt-out, deletion right or another region’s procedure also blocks model training.

A disclosure checklist before you send

1. Delete direct identifiers. Remove names, handles, addresses, phone numbers, private email addresses, account numbers and exact dates.
2. Broaden identifying combinations. Use an age range, larger geographic area, generic job description and approximate timeline when exact details add no value.
3. Keep protected material outside the chat. Do not paste passwords, authentication codes, payment credentials, medical records, legal correspondence or confidential workplace and school content.
4. Protect third parties. Substitute roles such as “my colleague” or “a relative,” then remove circumstances that could reveal the person.
5. Review every attachment separately. Crop unrelated people, documents and surroundings, and consider whether a voice recording itself reveals identity.
6. Use an available training control. Treat it as a restriction on the stated training use, not as a substitute for minimizing disclosure.

If you no longer want the account, Character.AI’s account-deletion instructions describe deletion as permanent and place the web command under Profile Settings, Account and Manage Account & Data; the mobile app uses Remove Account. Deletion is separate from deciding what to disclose: the strongest preventive boundary is still to submit only the information needed for the conversation.

Also read:

• Why It Is Easier to Succeed With An Exchange Rate API Than You Might Think
• Claude Projects Share More Than Chats—Check Knowledge and Permissions
• What is Business Model Canvas and How do You Create One