Connecticut Writes the Omnibus, Stanford Maps Delusional Spirals

2026-04-01

Conversational AI Watch

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

By Jess Jessop | April 1, 2026 | Issue #3

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

JESS'S TAKE

The White House told Congress last week to preempt state AI laws.

Ten days later, Washington's governor signed one of the strongest AI chatbot safety bills in the country.

That collision tells you everything about where this fight is heading.

Meanwhile, one senator in Connecticut is quietly doing something nobody else has tried. He wrote the AI bill AND the data privacy bill. Together. As a matched set.

No other state has both from the same author, moving at the same time.

And at Stanford, researchers just published the first study of real chat logs from people who experienced psychological harm from chatbot use. The finding: chatbots display sycophancy in more than 70% of their messages.

The people closest to the harm are moving fastest.

Here is what you need to know.

THE SIGNAL / REGULATION

Connecticut Is Writing the Omnibus Other States Will Copy

Senator James Maroney's SB 4 and SB 5 both advanced through the General Law Committee last week. SB 4 passed 16-4. SB 5 passed unanimously.

Most coverage treats them as separate bills. They are not.

SB 5 governs how AI systems behave: chatbot safety, employment AI, whistleblower protections for frontier developers, synthetic content disclosure. SB 4 governs what those systems can do with the data: broker registration, centralized deletion, algorithmic pricing disclosure, facial recognition restrictions, geolocation bans.

Together they form the most complete state-level framework for both AI behavior and data rights in the country.

CBIA, the state's business lobby, is pushing back on SB 4's broad data broker definitions. Their concern is legitimate. But the answer is tighter definitions, not weaker protections.

Takeaway: If your state is drafting AI or privacy legislation, study Connecticut's approach. Building AI governance and data rights as a matched set solves the problem every other state is creating by legislating them separately.

For Legislators: Connecticut's matched-set approach is a model worth stealing. AI governance without data rights is unenforceable. Data rights without AI governance is incomplete. Maroney built both, from the same committee. Ask your staff to pull SB 4 and SB 5 together.

CBIA · State of Surveillance

THE SIGNAL / AI SAFETY

Stanford Maps the Anatomy of a "Delusional Spiral"

Researchers at Stanford published the first in-depth analysis of real chat logs from 19 users who experienced psychological harm from AI chatbot interactions. The paper appears at ACM FAccT 2026.

The numbers are stark. Chatbots displayed sycophantic messages in more than 70% of their responses. More than 45% of all messages showed signs of delusions.

The most common pattern: the chatbot rephrasing what the user said, validating it, telling them their thoughts were unique and had grand implications. The chatbot was not grounding. It was amplifying.

The researchers identified acute cases where the chatbot encouraged self-harm or violent thoughts.

Takeaway: This is not hypothetical harm. This is documented harm from real users with real chat logs. Every chatbot safety bill in the country just got stronger evidence. And every AI mental health product without licensed clinical oversight just got a harder question to answer.

For Legislators: Cite this paper. It transforms chatbot safety from a theoretical concern into a documented clinical reality. The 70% sycophancy finding alone justifies mandated harm-detection protocols. Tech companies will tell you the next version will be safer. Stanford tested that claim and it is not true.

Stanford Spirals Project · Stanford HAI

THE SIGNAL / REGULATION

Washington Signs Chatbot Safety Law as White House Calls for Federal Preemption

Governor Bob Ferguson signed HB 2225 into law on March 24. Washington is now the second state with an enforceable chatbot safety law. Disclose the bot is not human at the start of every conversation. Repeat every three hours for adults. Every hour for minors. The law takes effect January 1, 2027.

Ten days earlier, the White House released a National Policy Framework for AI calling on Congress to preempt state laws that govern how AI models are developed. That directly targets legislation already enacted in California, Colorado, Texas, Illinois, and dozens of other states.

The collision is real. The federal framework prioritizes child safety but the preemption language puts it on a direct course against the state chatbot safety bills passing right now.

Takeaway: Federal preemption requires an act of Congress, and Congress has not passed comprehensive AI legislation in six years of trying. The states are not waiting. Washington just proved it.

For Legislators: Do not slow down. Continue building your state frameworks. If federal preemption eventually passes, your bills become the negotiating baseline for what the federal standard should include. Washington's HB 2225 does three things well: defines what an AI companion chatbot is, creates age-tiered protections, and mandates self-harm intervention protocols. Study it before writing from scratch.

Washington State Legislature · Holland & Knight

THE SIGNAL / CLINICAL AI

SDMH President Co-Authors Framework for Integrating Digital Mental Health Into Clinical Practice

Giovanni Ramos and Stephen Schueller (President, Society for Digital Mental Health) published a new paper in Cognitive and Behavioral Practice: "State of the Science in Clinical Practice: Digital Mental Health Interventions."

The paper moves past "do digital mental health interventions work?" into "here is how clinicians can identify, select, and integrate them into therapy today." That is a significant shift. The field has been stuck on efficacy debates for years. This paper gives clinicians actionable steps.

The equity lens is not an afterthought. Ramos focuses on digital mental health access for racially and ethnically minoritized populations. It is built into the framework.

Takeaway: When the president of the Society for Digital Mental Health publishes a practical integration framework, the field is signaling readiness. The question is no longer whether digital interventions belong in clinical practice. It is which ones, and how.

For Legislators: This paper provides the clinical evidence base for bills that create pathways for digital mental health tools, not just restrictions. Good legislation should require the safety architecture and clinical oversight, not ban the technology. The Ramos-Schueller framework shows what responsible integration looks like.

Center for Technology and Behavioral Health · UCI Psychological Science

THE SIGNAL / PATIENT SAFETY

ECRI Names AI Diagnostics the #1 Patient Safety Threat. The National Academy of Medicine Responds.

For the first time, an AI-specific concern holds the top position on ECRI's annual patient safety list. "Navigating the AI diagnostic dilemma" reflects the reality that AI tools can lead to missed, delayed, or incorrect diagnoses without proper safeguards and clinical oversight. Gaps and biases in training data can worsen health disparities.

Meanwhile, the National Academy of Medicine launched its "Patient Safety in the Era of AI" initiative on March 4. Co-chaired by the CEOs of Mayo Clinic and CommonSpirit Health, it builds on the landmark 2000 "To Err Is Human" report. This initiative will produce the recommendations that hospitals, health systems, and regulators use to set AI safety standards.

Takeaway: The most respected patient safety organization in the country says AI diagnostics is a bigger threat than workforce shortages and funding cuts. And the most prestigious medical academy in the country just mobilized its CEO-level leadership to write the rules. AI in healthcare is moving faster than safety infrastructure can keep up. If you are building AI for healthcare, these are the standards you will be measured against.

For Legislators: ECRI's list is used by hospitals, regulators, and insurers to prioritize safety spending. When you write AI healthcare bills, cite this list. The NAM initiative will likely produce the most authoritative set of AI patient safety recommendations in the country. Build legislative review triggers that update your statutes when NAM publishes its final recommendations.

Advisory Board · National Academy of Medicine

THE SIGNAL / MENTAL HEALTH POLICY

Raskin Introduces Suicide Prevention Bills as FDA Opens Behavioral Health Pathway

Representative Jamie Raskin introduced two bipartisan bills targeting the crisis infrastructure gap. The STOP Suicide Act creates a new SAMHSA grant program to expand access to outpatient and virtual stabilization care. The 9-8-8 Connect Act provides federal funding to the community organizations already answering 988 calls. These bills target the gap between "someone called for help" and "someone received sustained treatment."

At the same time, the FDA's TEMPO pilot, launched with CMS Innovation's ACCESS model, is creating the first joint FDA-CMS reimbursement-linked pathway for digital mental health devices. The pilot covers four clinical tracks, including behavioral health targeting depression.

Takeaway: Two signals moving in the same direction. Congress is funding the crisis infrastructure that chatbot safety bills will point clients toward. And the FDA is building a reimbursement pathway for digital mental health devices that meet safety standards. The pipeline from crisis detection to sustained treatment to reimbursement is being built right now.

For Legislators: State chatbot safety laws create the referral mandate. Raskin's bills fund what happens after the referral. And TEMPO creates the reimbursement pathway for digital tools that meet FDA standards. If your state is writing AI mental health bills, reference all three. Together they form a complete pipeline from detection to treatment to payment.

Rep. Raskin · Manatt · Wilson Sonsini

WHAT WE BUILT

Casey AI: The Therapist in the Loop Standard

Casey is our AI behavioral health assistant built on the architecture the market is validating: clinician supervision at every layer. 1.78 million safety tests. 100% accuracy on crisis detection. Pursuing FDA 510(k) clearance.

Casey does not improvise therapy. Casey does not replace clinicians. Casey makes clinicians more effective by extending their reach between sessions, under their direct oversight.

Watch the Demo →