Chatbots face new safety rules

Summary#

HB 635 would create the Artificial Intelligence Chatbots Act and amend Virginia’s Consumer Protection Act (VCPA). Its stated goal, as reflected in the bill’s title, is to establish rules for AI chatbots and prohibit specified practices. It is not law; the bill says it would take effect January 1, 2027.

  • The rules would cover “companion chatbots” made available in Virginia: generative AI systems that simulate an ongoing, human-like relationship. The bill excludes several customer-service, workplace, business, and video-game bots, and AI systems used only for research before being made available.
  • Operators could not make a companion chatbot available to a minor if it is capable of specified harmful conduct, including encouraging self-harm, offering unsupervised mental-health therapy, or engaging in sexual interactions with minors.
  • Operators would have to use commercially reasonable methods to determine users’ ages, give persistent and timed notices that the chatbot is not human, and avoid implying that its output comes from a licensed professional.
  • Operators would have to use a protocol to detect expressions of suicidal thoughts or self-harm and refer users to crisis services; get a parent or guardian’s written consent before using a minor’s personal information to train the chatbot’s underlying model; and publish specified safety information and reports.
  • Violations would become prohibited practices under the VCPA and be subject to its enforcement provisions. The bill creates a new chapter of law, not a new agency.

What it means for you#

  • Chatbot operators: If you make a covered companion chatbot available in Virginia, you would have to meet the age, safety, notice, crisis-response, privacy, and public-reporting requirements. The bill does not set out an application or approval process.
  • Minors and parents or guardians: A covered chatbot could not be made available to a minor if it is capable of the listed harmful conduct. A parent or guardian’s affirmative written consent would be needed to use a minor’s personal information to train the underlying model.
  • Other users: Operators would have to show that the chatbot is not human and provide crisis-service referrals when the protocol detects expressions of suicidal thoughts or self-harm. The bill does not require a chatbot to provide therapy.
  • Users and the public: Operators would have to provide a way to report adverse incidents, publish an anonymized and aggregated incident catalog, publish safety-test findings, and issue public reports every six months.
  • People using other AI tools: The bill does not cover every AI system. Its rules apply to the defined category of companion chatbots, subject to the listed exclusions.

Money#

No publicly available information on estimated public or private costs.

  • The bill requires operators to take steps such as age screening, safety testing, reporting, and crisis referrals, but gives no cost estimate.
  • It does not state an appropriation, fee, or penalty amount. It makes violations subject to VCPA enforcement provisions; the supplied material does not include the current VCPA text or its remedies.

What is unclear#

  • The bill does not define “commercially reasonable,” “reasonable efforts,” or when a chatbot is “capable” of prohibited conduct.
  • It does not set detailed rules for age screening, safety tests, incident reports, or oversight of compliance.
  • The age rule refers to what happens before January 1, 2027, but the bill itself would take effect on that date. The practical role of the earlier date is unclear.
  • The supplied material does not include the current VCPA text, so I could not verify which enforcement provisions already apply or exactly how the new prohibited practice would change them.

Case for#

  • A possible argument for the bill is that it sets specific safeguards for minors interacting with chatbots designed to simulate ongoing personal relationships.
  • Requiring crisis referrals when a chatbot detects expressions of suicidal thoughts or self-harm could give users a route to human crisis services.
  • Persistent notices and limits on claims about licensed professionals could help users understand they are interacting with a chatbot, not a person or licensed professional.
  • Public incident and safety information could make some risks easier to identify.

Case against#

  • The bill’s key standards are not fully defined. Operators may face uncertainty about what counts as reasonable safeguards or prohibited capability.
  • It requires several systems and public reports but does not set detailed procedures or oversight rules in the supplied text.
  • A chatbot’s ability to produce harmful content could be difficult to assess; the bill does not explain how that capability would be tested or judged.
  • The bill does not state costs or penalty amounts, and the supplied material does not show how the new rules would interact with existing VCPA enforcement.