JAMA Psychiatry: ChatGPT Reinforces Psychotic Thinking. The Free Version Is Worst.

2026-03-30

Forty million Americans used ChatGPT for healthcare advice last year. Zero of those conversations had a licensed clinician in the loop. A college student took her own life after months confiding in an AI chatbot named Harry. Families are now testifying in state legislatures. Juries are returning verdicts against tech companies. And the regulatory wave that everyone said was years away? It arrived this quarter.

Conversational AI Watch exists because the people making investment decisions, building products, writing laws, and setting policy need to see these signals before they become headlines. Every morning, I scan the intersection of AI safety, mental health, regulation, and care delivery and surface what moves money and changes law.

Here is what you need to know right now.

THE SIGNAL / BREAKING RESEARCH

JAMA Psychiatry: ChatGPT Reinforces Psychotic Thinking Across All Versions

Published March 25 in JAMA Psychiatry, Columbia University psychiatrists tested all three versions of ChatGPT with 79 statements a person experiencing psychosis could plausibly make. The free version had 43-fold higher cumulative odds of giving inappropriate responses to psychotic prompts. Even GPT-5 generated inappropriate responses at nine times the control rate.

The researchers' conclusion: the free version most likely to affirm unstable thinking is the one used by economically disadvantaged users, who are already at greater risk for psychosis.

Takeaway: This is JAMA Psychiatry, the gold standard, saying the most widely used AI chatbot on earth actively reinforces psychotic thinking. Without a clinician in the loop, there is no safety net.

For Legislators: The most popular AI chatbot in the world makes psychotic people worse, and the free version that low-income users rely on is the most dangerous. Any bill regulating AI chatbots should require independent safety testing against vulnerable populations before these products are allowed to operate, not after people are harmed.

Sources: JAMA Psychiatry · Washington Times

THE SIGNAL / REGULATION

98 Chatbot Bills Across 34 States. The Regulatory Wave Is Here.

The Future of Privacy Forum is tracking 98 chatbot-specific bills across 34 states and three federal proposals in 2026. Oregon and Washington have given final approval to safety laws. Illinois and Nevada banned AI behavioral health services outright. Kentucky, Missouri, Vermont, and North Carolina all have bills regulating AI in therapy.

Takeaway: If your AI product talks to people about how they feel, a licensed professional needs to be structurally integrated into the architecture. Not bolted on. Built in.

For Legislators: Ninety-eight bills means ninety-eight different definitions of what "safe" means. The industry needs one clear standard, not a patchwork. Consider referencing an independent safety certification in your bills, the same way healthcare law references Joint Commission accreditation. Write the legal framework. Let an independent body write the technical one.

Sources: Future of Privacy Forum · Troutman Pepper

THE SIGNAL / GLOBAL

WHO Declares All Conversational AI a Public Mental Health Concern

On March 20, WHO published formal recommendations calling for generative AI use to be recognized as a public mental health concern. The key phrase: this applies to all generative AI solutions, not only those intended for mental health. WHO is establishing a Consortium of Collaborating Centres on AI for Health across all six global regions.

Takeaway: When WHO says "all generative AI," they mean ChatGPT, Claude, Gemini, every companion chatbot, every customer service bot that asks "how are you doing today?" The regulatory perimeter just expanded beyond anything the industry expected.

For Legislators: The World Health Organization just said this is not just a therapy chatbot problem. It is an all-AI problem. Any chatbot that has a personal conversation with a person can affect their mental health. Your bills should not be limited to apps that call themselves "mental health tools." They should cover any AI that holds ongoing personal conversations with people.

Sources: WHO

THE SIGNAL / FUNDING

Grow Therapy Hits $3B. Limbic Publishes in Nature Medicine.

Sequoia-backed Grow Therapy closed $150M at a $3B valuation. Spring Health's acquisition of Alma created a combined $688M entity. Meanwhile, Limbic published in Nature Medicine showing AI agents with a clinical reasoning layer outperformed licensed human therapists on CBT scoring. The catch: it works because of the clinical architecture layer, not because general-purpose LLMs are therapeutic.

Takeaway: The market rewards two things: companies putting licensed professionals at the center, and companies proving clinical architecture changes outcomes. General-purpose chatbots doing therapy without either are getting legislated out of existence.

For Legislators: AI can actually help people with mental health when it is built right. The Nature Medicine study proves that. The difference between AI that helps and AI that harms is whether a clinical safety layer and licensed professional oversight are built into the system. Good legislation should require the safety architecture, not ban the technology.

Sources: Bloomberg · Limbic / Nature Medicine

THE SIGNAL / RESEARCH

Brown and Stanford Prove Unsupervised Chatbots Systematically Fail

Brown University identified 15 distinct ethical risks when AI chatbots are prompted to act as therapists, including "deceptive empathy" that mimics care without understanding. Stanford HAI found persistent stigma toward alcohol dependence and schizophrenia, and that newer, larger models showed no improvement over older ones.

Takeaway: Two top-tier institutions, same conclusion. Prompting a general-purpose LLM to "act like a therapist" does not make it safe. Bigger models do not fix broken architecture.

For Legislators: Tech companies will tell you "the next version will be safer." Brown and Stanford tested that claim and it is not true. Newer, bigger, more expensive AI models are just as dangerous to vulnerable people as older ones. Safety does not come from better AI. It comes from putting a licensed human professional in the system who is accountable when something goes wrong.

Sources: Brown / ScienceDaily · Stanford HAI

JESS'S TAKE

Count the converging signals in just one week: JAMA Psychiatry proves ChatGPT reinforces psychotic thinking. Brown identifies 15 ethical failures. Stanford shows bigger models do not fix stigma. Limbic proves clinical architecture changes outcomes. WHO says all conversational AI is a public health concern. Thirty-four states are writing laws. And the market is sending $3 billion to companies that put licensed professionals at the center.

The era of "move fast, ship a chatbot, figure out safety later" is over. The companies that will own the next decade of mental health AI are the ones building licensed clinician oversight into the architecture from day one. Not as a feature. As the foundation.

That is what we are building at BetterMind. That is what Therapist in the Loop means. And the window for companies that have not figured this out is closing fast.

WHAT WE'RE BUILDING

Casey AI just completed 1.78 million pre-FDA safety test executions at 100% accuracy. Every response supervised by a licensed therapist. That is the architecture the research, the regulation, and the market are all demanding.

See Casey in action