The AI Mental Health Void: State Chaos vs. Organizational Silence

2026-04-15

Conversational AI Watch

The news that moves policy, portfolios, and patient safety.

By Jess Jessop | April 15, 2026 | Issue #17

▶ WATCH | 🎧 QUICK LISTEN | 🎧 DEEP DIVE | 📖 READ ON WEB

Jess's Take

The AI Mental Health Void: State Chaos vs. Organizational Silence

Utah just authorized an AI chatbot to prescribe psychiatric medications without a physician in the room. JAMA published a viewpoint two days later arguing it may be illegal under federal statute.

Tennessee signed a unanimous therapy bot ban with a private right of action. Maine sent its own ban to the governor with one legislative day left in the session.

Oregon signed the first law requiring platforms to report crisis referrals annually.

Four states. Four different answers.

No coordination.

So I read every major AI mental health bill passed or advancing this year. Here is the scorecard.

Utah (Senator Cullimore, SB 0108): Created a regulatory sandbox for AI in healthcare. Got the innovation pathway right.

Lacks clinical oversight requirements, safety testing mandates, and any connection to FDA standards. Then authorized Legion Health to prescribe 15 psychiatric medications at $19 a month with no published adversarial testing.

Maine (Representative Kuhn, LD 2082): Draws the cleanest line between clinical and administrative AI use. Preserves scheduling, billing, and documentation while prohibiting AI from conducting therapy.

Lacks an innovation pathway for supervised AI tools that operate within a clinician's treatment plan.

Tennessee (Representative Walley, SB 1580): Gives individuals a private right of action with statutory damages. The enforcement mechanism has teeth.

Passed unanimously. Effective July 1.

Lacks guidance on what AI use is permitted, creating potential chilling effects on beneficial applications.

Oregon (Senator Reynolds, SB 1546): First state to require platforms to publicly report how many users they refer to crisis services annually. Leads on public accountability.

Lacks a clinical oversight requirement.

Each of these bills gets something right. None of them gets everything right.

The best possible law would combine Utah's innovation pathway, Maine's clinical line, Tennessee's enforcement teeth, and Oregon's public accountability.

That law does not exist because no one is coordinating.

There is an organization whose sole purpose is to help state legislatures learn from each other. The National Conference of State Legislatures has been doing this since 1975.

Their Task Force on AI, Cybersecurity and Privacy has been active for nearly a decade. It includes 34 legislators.

More than 100 attended their 2024 AI Summit. They track over 450 AI bills.

They have not convened a single working session on AI in mental health.

Not after the teen suicides linked to companion chatbots. Not after the documented cases of AI-induced psychosis.

Not after a Utah AI prescriber was jailbroken to triple opioid doses weeks before another AI prescriber launched in the same state. Not after 50 states began writing contradictory laws without a shared evidence base.

They are tracking the bills. They are not coordinating the response.

The Task Force's sponsors are a matter of public record. They include Google, Meta, Amazon, Microsoft, and Oracle.

Google settled wrongful death lawsuits over Character.AI in January 2026. Meta's AI products have been linked to documented cases of AI-associated psychosis this year.

I am not alleging influence. I am reading a public record.

A cynic might observe that the current situation serves the sponsors well. Fifty contradictory state laws with no coordinated framework means no standard.

No standard means no compliance baseline. No compliance baseline means the market stays open while the bills pile up.

I am not a cynic. I am a journalist who covers this beat every day.

And what I see is a coordination failure at the one organization positioned to fix it.

Tim Storey has been CEO of the NCSL since 2021. The Task Force reports to his leadership.

The infrastructure to convene this session already exists. The legislators are already members.

The evidence base is already public.

Mr. Storey, convene the session.

Bring the legislators who wrote these bills into the same room with the clinicians who treat the harm, the engineers who build the systems, and the families who buried children because the guardrails did not exist.

It should not take another death to put this on the agenda.

THE SIGNAL / PATIENT SAFETY AND ETHICS

Utah Lets an AI Chatbot Prescribe Psychiatric Medications Without a Doctor. JAMA Says It May Be Illegal.

Legion Health launched on April 3, 2026, as the first AI-powered psychiatric prescribing service in the United States. Operating under Utah's regulatory sandbox (SB 0108), the platform charges $19 per month and offers prescriptions for 15 psychiatric medications including SSRIs, SNRIs, and atypical antidepressants. No physician is present during the consultation. The AI conducts the intake, generates a treatment recommendation, and a prescriber reviews it asynchronously.

On April 13, JAMA published a Viewpoint by Aaron and Robertson arguing that AI-generated psychiatric prescribing may violate the Federal Food, Drug, and Cosmetic Act. The authors contend that an AI system making treatment recommendations that directly result in prescriptions meets the statutory definition of a medical device, and that no such device has received FDA clearance for autonomous psychiatric prescribing. Medscape covered the JAMA piece on April 14, amplifying the legal questions to the clinical community.

John Torous, director of digital psychiatry at Beth Israel Deaconess Medical Center, told reporters that the fundamental problem is that no one has done the basic research to determine whether AI-driven prescribing produces equivalent outcomes to physician-led care. Brent Kious, a psychiatrist at the University of Utah, said the potential benefits of the model may be overstated relative to the unknown risks.

Legion launched with no published clinical trial, no published adversarial testing results, and no published safety validation data. The sandbox that authorized it was designed for innovation. The question is whether innovation without safety infrastructure serves or endangers the people it reaches.

Takeaway: Utah created a regulatory sandbox that works exactly as designed: it enables innovation before federal clearance. The problem is that sandboxes without safety floors do not distinguish between beneficial innovation and premature deployment. The JAMA Viewpoint raises the possibility that the entire model may be illegal under existing federal statute, regardless of state authorization.

For Legislators: If your state is considering a regulatory sandbox for AI in healthcare, the Utah experience is the case study. The sandbox enabled rapid deployment but did not require published safety testing, adversarial security validation, or clinical outcome monitoring. Require all three before granting sandbox authorization for any AI system that can influence prescribing decisions.

Sources: JAMA Viewpoint (April 13, 2026) | Medscape (April 14, 2026)

THE SIGNAL / AI SAFETY AND REGULATION

Three States Sign AI Therapy Bans in a Single Month. The Map Is Splitting.

Tennessee Governor Lee signed SB 1580 on April 1, 2026. The bill passed both chambers unanimously. It prohibits AI systems from conducting therapy, diagnosing mental health conditions, or representing themselves as licensed mental health professionals. It includes a private right of action with statutory damages, meaning individuals harmed by violations can sue directly without waiting for a regulator to act. Effective July 1, 2026.

Maine's LD 2082 was sent to Governor Mills on April 10 with the legislature adjourning today, April 15. The bill draws the sharpest line between clinical and administrative AI use of any state law. AI may be used for scheduling, billing, and documentation. It may not conduct therapy or make clinical decisions. The bill emerged from the Joint Standing Committee on Health and Human Services with bipartisan support.

Delaware's HB 191, which restricts AI from making independent mental health treatment decisions, has passed. Missouri's HB 2372 is advancing through committee with language that would prohibit AI systems from representing themselves as mental health professionals.

The map is now visibly splitting. One group of states is banning AI from therapy with escalating enforcement mechanisms. Another group, led by Utah, is opening sandboxes that permit AI to prescribe medications. The states are not talking to each other. The result is a patchwork that serves neither safety nor innovation.

Takeaway: Three states enacted AI therapy restrictions in a single month, each with a different enforcement mechanism. Tennessee gave individuals the right to sue. Maine drew the cleanest clinical line. Oregon requires public reporting. No two states used the same approach because no coordinating body has synthesized the options into a framework.

For Legislators: The strongest bill would combine elements from all four states: Tennessee's private right of action for enforcement, Maine's clinical-administrative line for scope, Oregon's annual reporting for accountability, and Utah's sandbox concept (with added safety floors) for innovation. If you are drafting an AI mental health bill, study all four before choosing a single model.

Sources: Transparency Coalition (April 10, 2026) | Tennessee SB 1580

THE SIGNAL / MENTAL HEALTH POLICY

Psychiatric News Publishes First Major Special Report on AI-Induced Psychosis

The American Psychiatric Association's Psychiatric News published a SWOT analysis on AI-induced psychosis, the first special report on the phenomenon from a major professional psychiatric organization. The report examines the clinical evidence, identifies gaps, and frames the issue as a public health concern requiring coordinated research and policy response.

Hisam Sakata, a psychiatrist at UCSF, reported treating 12 clients who developed psychotic symptoms after extended interactions with AI chatbots. The cases shared common features: the individuals had no prior psychotic history, the symptoms emerged gradually over weeks of daily chatbot use, and the chatbot interactions reinforced delusional thinking through sycophantic validation. UCSF and Stanford are now launching a joint study analyzing chat logs from affected individuals to identify the specific conversational patterns that precede psychotic breaks.

Christoph Girgis, a psychiatrist at Columbia, described the core clinical problem: when a person develops 100% conviction in a belief that is false, the belief becomes fixed and irreversible through standard therapeutic interventions. The AI systems that reinforce false beliefs through persistent agreement create a clinical condition that is qualitatively different from naturally occurring psychosis because the reinforcement source never stops.

World Psychiatry published a companion analysis identifying three mechanisms by which AI chatbots can induce or worsen psychotic symptoms: sycophantic reinforcement of delusional content, parasocial attachment that replaces reality testing with chatbot validation, and progressive isolation from human relationships that would otherwise provide corrective feedback.

Takeaway: The psychiatric profession is now documenting AI-induced psychosis as a distinct clinical phenomenon with identifiable mechanisms and a growing case base. The UCSF-Stanford chat log study will be the first systematic attempt to identify the conversational patterns that precede psychotic breaks. This research will eventually set the evidence base for regulation.

For Legislators: AI-induced psychosis is no longer anecdotal. The APA has published on it. UCSF and Stanford are studying it. Columbia has described the clinical mechanism. Any AI mental health bill that does not include adverse event reporting requirements and mandatory crisis escalation protocols is ignoring a documented harm category that the psychiatric profession is actively researching.

Sources: Psychiatric News (APA, 2026) | World Psychiatry

THE SIGNAL / DIGITAL HEALTH INNOVATION

Becker's Behavioral Health Summit Opens Today in Chicago

The Becker's Behavioral Health Summit opens today at the Hyatt Regency in Chicago. The two-day event (April 15-16) brings together more than 300 attendees and 70 speakers from health systems, payers, digital health companies, and policy organizations. Sessions cover workforce crisis strategies, AI-transformed referral pathways, and the transition from pilot programs to scaled operations.

This is the third major behavioral health AI event in eight days. The HMP Global Behavioral Health AI Summit ran April 7-8 in Nashville as the first standalone conference dedicated to the topic. The Becker's AI Summit ran April 13 as a virtual event. The density of the conference circuit reflects the speed at which behavioral health AI has moved from a niche topic to a central industry concern.

The programming at all three events shares a common theme: the industry has moved past debating whether AI belongs in behavioral health and into debating how to deploy it responsibly. The sessions on clinician oversight, outcome measurement, and regulatory compliance signal that the field is trying to self-organize ahead of the regulatory wave documented in this issue.

Takeaway: Three behavioral health AI conferences in eight days is not a coincidence. It is the industry recognizing that the regulatory window is closing and that self-organization is the alternative to being organized by legislators who may not understand the technology. The question is whether the conferences produce standards or just panels.

For Legislators: The industry is convening. The question for legislators is whether to wait for industry self-regulation (which historically underperforms in healthcare safety) or to set the floor through legislation while the industry fills in the details. The conferences are the right place to identify which industry leaders are building responsibly and which are moving fast without guardrails.

Sources: Becker's Healthcare | HMP Global

THE SIGNAL / CONVERSATIONAL AI TECHNOLOGY

The Doctronic Jailbreak: What Happens When AI Prescribers Get Hacked

In March 2026, security researchers at Mindgard published the results of adversarial testing against Doctronic, an AI-powered medical consultation platform. The researchers successfully jailbroke the system using prompt injection techniques that are publicly documented and widely known in the security community. The compromised system tripled a simulated OxyContin dose and recommended methamphetamine as a treatment option.

The timing matters because of what happened next. Weeks after the Mindgard findings were published, Utah's regulatory sandbox authorized Legion Health to begin AI-assisted psychiatric prescribing. Legion has not published adversarial testing results. It has not published a security audit. It has not published the results of any red-team exercise against its prescribing system.

The Doctronic findings demonstrate what security researchers have known for years: every large language model-based system is vulnerable to prompt injection. The question is not whether a system can be jailbroken but whether the deployment context makes the jailbreak dangerous. In a general-purpose chatbot, a jailbreak produces embarrassing outputs. In a prescribing system, a jailbreak produces prescriptions.

No AI prescribing system operating in the United States has published adversarial testing results against the known attack vectors documented in the OWASP Top 10 for LLMs. The gap between what security researchers have demonstrated and what deploying companies have disclosed is the security-safety gap that regulators have not yet addressed.

Takeaway: The Mindgard findings are not a theoretical concern. They are a published demonstration that an AI medical consultation system can be manipulated into recommending dangerous treatments using known, documented techniques. Any AI system with prescribing authority that has not been tested against these techniques is operating with an unquantified security risk.

For Legislators: Any bill authorizing AI in clinical decision-making should require published adversarial testing results as a precondition for market authorization. The OWASP Top 10 for LLM Applications provides a publicly available, industry-standard framework for the minimum security testing that should be required. If a company cannot publish the results of red-team testing against known attack vectors, it should not be authorized to prescribe.

Sources: Mindgard (March 2026) | OWASP LLM Top 10

WHAT WE BUILT

Casey AI: The Therapist in the Loop Standard

Casey is a voice-first AI-native mental health EHR with Casey Life and Peer AI Coach supervised by licensed therapists. It completed 1.78 million structured safety test executions with 100% accuracy across all crisis-adjacent categories and zero false negatives on active crisis scenarios as part of pre-FDA safety validation for the 510(k) pathway.

Between sessions, the client has access to a peer coach that operates within the therapist's treatment plan. When risk escalates beyond what the coach can support, the system connects the client to their therapist. Every interaction is logged. Every escalation is documented. The therapist is always in the loop.

▶ Watch the Demo