Virginia proposal would set safety rules for companion chatbots

Operators would have to keep a companion chatbot from being available to a minor if it can engage in listed unsafe conduct. A minor is someone younger than 18. A companion chatbot is a conversational AI system designed to simulate an ongoing human-like relationship; certain customer service and other systems are excluded. Operators would have to use commercially reasonable methods, such as a neutral age screen, to determine whether a user is a minor. Before January 1, 2027, this restriction would not apply if the operator lacked actual knowledge that the user was a minor. Starting that date, it would not apply if the operator reasonably determined the user was not a minor. The act would take effect January 1, 2027. The rule for deciding whether a user is a minor would change on that date.

Operators would have to tell users that a companion chatbot is not human and provide a way to reach crisis services when it detects expressions of self-harm or suicidal thoughts. These requirements cover operators of companion chatbots and users who interact with them. The chatbot would have to show a lasting notice and notify users at login, every 90 minutes of sustained engagement, and when prompted. If a pop-up is not feasible, another form of communication could be used. Chatbot advertising and interfaces could not imply that outputs come from a licensed professional. The act would take effect January 1, 2027. Notices would be required at login, every 90 minutes of sustained engagement, and when requested.

Operators could not use a minor's inputs to train a companion chatbot's underlying model without a parent's or guardian's written consent. This rule covers a minor's inputs to a companion chatbot and the minor's personal information. The parent or guardian would have to affirmatively give written consent to use the minor's personal information for that specific purpose. The act would take effect January 1, 2027.

Operators would have to collect user reports of chatbot-related harmful incidents and publish incident information, safety test findings, and reports on specified chatbot responses and crisis referrals. These duties would apply to operators of companion chatbots. Incident information would have to be anonymized and grouped together before publication. Operators would have to publish safety test findings for tests conducted to meet the minor-safety rules. Public reports would count responses involving suicide, self-harm, suicidal thoughts, harming others, or illegal activity, and times a mental health referral was provided. The act would take effect January 1, 2027. The report on chatbot responses and referrals would be published twice a year.

Violations of the chapter would count as prohibited consumer practices and could be enforced under Virginia's Consumer Protection Act. This would apply to violations of the proposed Artificial Intelligence Chatbots Act. The act would take effect January 1, 2027.