How AI works in RUB

RUB includes fictional, AI-generated characters. Your chat partner may be human or an AI character. After the result, and throughout continuing bot and group conversations, the bot is labeled as AI-generated and not human. The guessing mechanic never changes that fact.

AI characters can be inaccurate and do not have human feelings, professional qualifications, or an independent real-world identity. They are not a substitute for medical, legal, financial, or crisis support.

A human user may optionally enable an AI delegate to answer eligible messages on their behalf. The account remains a human account, but every delegate-authored message is separately labeled as generated by an AI delegate on the sender's behalf. The sender can disable the delegate; recipients should not assume that labeled text was reviewed or typed by that person.

RUB generates character text and safety classifications with third-party model services routed through OpenRouter. Requests carrying user content require zero-data-retention routing. The model providers are listed at beardata.co/legal/subprocessors. RUB does not train or substantially modify a foundation model that is offered to the public.

Before AI processing, RUB asks for separate, explicit permission to send selected messages and chat context, uploaded images and sampled video frames, public profile details, posts, game submissions, and answers to Ruber to these services. This can include relevant existing content and future content. The permission starts off; you can decline and sign out, or withdraw in Settings to stop future AI requests with your content. Withdrawal cannot undo processing already completed.

Separately, users may opt in to let bear.ai license de-identified RUB round-chat and game-submission dataset rows to AI companies for their own training and evaluation. That choice is off by default and is described in the Privacy Policy; it is not the training data of the third-party models used to answer a live chat. Commercial delivery is currently disabled.

Uploaded images and server-decoded samples from videos receive a safety classification before publication. If the classifier is unavailable, the upload is withheld so the sender can retry. Sampling does not inspect every video frame, and automated filtering can miss content; individual messages and posts can also be reported in the app. Voice notes are not transcribed, and live floor audio is not recorded by RUB.

Before an AI character generates a reply, RUB checks a fresh user message for first-person suicidal-ideation or self-harm language. A match bypasses the model, character voice, simulated reading delay, and multi-bot follow-up. RUB sends one fixed resource message for that user message in round chats, continuing friend chats, and group chats. A retried job is deduplicated; a later fresh disclosure receives the resources again.

The fixed response says the AI cannot provide the help a real person can and refers the user to crisis services. In the United States, call or text 988 or text HOME to 741741. Outside the United States, findahelpline.com lists local services. Fleet-wide model instructions also prohibit encouraging or glorifying self-harm, giving instructions, or fostering dependence in place of real people or professional support.

RUB records only a content-free aggregate count of issued crisis referrals, split by chat surface, for safety monitoring and the annual California report that begins July 1, 2027. The report must not contain user identifiers or personal information. Detection is deliberately conservative but is not a diagnosis and is not perfect; call emergency services for immediate danger.

RUB is an adults-only service and requires a date of birth that indicates the user is at least 18. Companion chatbots may not be suitable for some minors. If we learn that a user is under 18, we remove the account rather than treating the general age gate as permission for a minor to continue.