Character.AI's training opt-out is limited to two regions

The design

Character.AI publishes a dedicated Training Data Documentation article, issued under a California disclosure law, stating that its generative AI systems use “publicly available data; text content and interaction data from users; internally generated data from safety training, engineering testing, and red-teaming exercises; synthetic data… and; third parties through commercial agreements.” The document says Character.AI “started collecting and using data for model development in 2021” and continues to use large-scale data for what it calls “post-training,” a process of customizing and fine-tuning existing models.

What the evidence says

A separate support article, How do I manage/update my model training settings?, names an actual control: a “Data & Privacy” or “Model Improvement” setting with a toggle called “Improve the Model for Everyone.” The article states this opt-out is offered specifically “If you are in the European Economic Area (EEA) or the United Kingdom (UK)” — it does not describe an equivalent toggle for users elsewhere, including the United States. Character.AI's own Regional Privacy Disclosures separately describe a broader-sounding “right to opt out at any time,” but point to this same EEA/UK-labelled article for the mechanism.

What it asks of people

A user outside the EEA and UK who wants to stop their conversations feeding model training has, on this documentation's own terms, no named self-service control. For users who can use the toggle, opting out is described as forward-looking only: “new content will not be used to train our generative AI models,” with no statement about removing previously used content from a model already trained on it. The article also notes that opting out does not stop Character.AI from using data “to improve other aspects of” the service, such as “search, recommendations, and safety classifiers.”

Privacy and safeguards

The training-data article states the company takes “steps to reduce the amount of personal information” in training data and to de-identify it before use, but neither article describes that process in technical detail or names an independent check on it. Neither document states whether the EEA/UK restriction reflects a legal requirement specific to those regions or a design choice Character.AI could extend elsewhere.

• Why is the named opt-out control limited to the EEA and UK rather than offered everywhere?
• Can a user confirm whether a specific past conversation was used in a completed round of post-training?
• What, precisely, changes for a user who opts out, beyond “new content will not be used”?

Character.AI's documentation is more specific than many companion-app policies about where its training data comes from and what opting out does and does not cover, and that same specificity is what makes the regional limit on the opt-out control visible rather than hidden in vaguer language.