August 23, 2026
Rhode Island enacted two laws on June 23 that regulate artificial-intelligence chatbots and their use in mental health care. One requires developers to build protocols for users who express suicidal ideation, self-harm, or threats of harm. The other addresses AI use in mental health treatment and reflects concern that automated systems can influence clinical decisions without the judgment, accountability, or relationship expected of licensed professionals.
Rhode Island is not alone. Florida legislation effective July 1, 2026 restricts AI use by clinical social workers, marriage and family therapists, and mental health counselors to administrative or supplementary functions, with specific conditions for recording or transcribing sessions. Washington has enacted rules directed at AI companion chatbots, emphasizing psychological safety, transparency, and the risks of systems that imitate empathy and intimacy. A New Mexico court’s August order against Meta added chatbot restrictions and warnings for minors.
The emerging pattern is a state-by-state safety regime built faster than social work institutions are revising practice standards.
The wrong binary
Professional debate often treats AI as either innovation or replacement. That binary is too crude. A scheduling tool, a transcription service, a risk-score generator, a therapeutic chatbot, and an autonomous crisis response system present different risks. Regulation should be based on function, data, and consequence.
Sociotechnical-systems theory holds that outcomes arise from the interaction of technology, institutions, workers, rules, and users. Harm is rarely located in an algorithm alone. A chatbot may provide a dangerous answer, but organizational decisions determine whether it was marketed as therapy, whether clinicians reviewed it, whether crisis escalation existed, and whether users understood they were speaking with software.
Street-level bureaucracy adds another insight. Human-service workers already exercise discretion under time pressure and resource constraints. AI can support that discretion, narrow it, or obscure it. When an automated recommendation appears authoritative, workers may defer even when context contradicts it—a phenomenon known as automation bias.
What social work needs now
First, every agency should maintain an AI inventory. Leadership cannot govern tools it has not identified. The inventory should include vendor, function, users, data inputs, outputs, client population, decision consequence, and whether the system communicates directly with clients.
Second, agencies should classify uses by risk. Scheduling and spelling assistance are low-risk. Drafting clinical notes, predicting suicide, recommending placement, assessing child safety, or communicating with a distressed minor are high-risk. High-risk uses require formal approval, testing, human review, and an exit plan.
Third, informed consent must be specific. A general technology clause is insufficient. Clients should know when AI records, transcribes, summarizes, recommends, or communicates; what data leave the agency; how long data are retained; and whether refusal affects service.
Fourth, social workers must remain accountable for professional judgments. “The system recommended it” is not an ethical defense. Agencies should prohibit automatic adverse decisions in child welfare, housing, benefits, involuntary treatment, and service termination.
Fifth, crisis protocols must be tested rather than assumed. Developers should demonstrate how a chatbot recognizes ambiguous statements, handles language differences, contacts human help, preserves privacy, and avoids false reassurance. Agencies should run simulations before exposing clients.
Sixth, procurement contracts should include audit rights, incident reporting, data deletion, model-change notification, bias testing, and indemnification. A vendor’s trade-secret claim cannot override a client’s right to understand a consequential decision.
Seventh, schools should teach AI governance across the curriculum. Students need more than prompt-writing. They need privacy law, human-subjects ethics, algorithmic bias, validation, procurement, documentation, and organizational change.
Eighth, professional boards should issue function-based guidance. Blanket bans can drive tools underground; vague encouragement leaves clients unprotected. Rhode Island’s approach should be evaluated through public incident data, enforcement reports, and independent research.
A profession-wide reform
Social work’s comparative advantage is not that humans are always wiser than machines. It is that professional practice is embedded in relationship, context, duty, and accountability. Technology can assist those commitments only when governance makes them operational.
The new state laws are early boundary markers. They should prompt NASW, CSWE, licensing boards, employers, and service users to create a common framework before 50 incompatible regimes define practice by accident.
Selected sources
Rhode Island General Assembly, June 23, 2026 enactment announcement:
https://www.rilegislature.gov/pressrelease/_layouts/15/ril.pressrelease.inputform/DisplayForm.aspx?ID=376765&List=c8baae31-3c10-431c-8dcd-9dbbe21ce3e9
Florida Senate Bill 344 (2026):
https://www.flsenate.gov/Session/Bill/2026/344/BillText/Filed/HTML
Washington AI companion chatbot statute:
https://app.leg.wa.gov/RCW/default.aspx?cite=19.440.010
California behavioral-sciences board analysis of chatbot regulation:
https://www.bbs.ca.gov/pdf/agen_notice/2026/20260507_item28_b.pdf