The Paper Trail

2026-04-18

Conversational AI Watch

Issue #20 • April 18, 2026 • By Jess Jessop

AI safety, mental health policy, and patient safety at the intersection of conversational AI

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

Jess's Take

The Paper Trail

This week, I kept finding the same word in five different stories.

Documented.

A law firm tracked 240 AI bills across 43 states and documented the pattern. The White House tried to kill them. Fifty Republican lawmakers documented their objection to their own president. The bills kept coming.

NPR sent a reporter to Kaiser Permanente and documented what happened after the strike we covered in Issue 13. The triage team went from nine to three. Licensed clinicians replaced by unlicensed operators reading scripts.

A team at the University of Montreal collected 71 media reports of people who experienced psychiatric harm after interacting with AI chatbots. They mapped all 36 unique cases. Then they checked whether any had been verified through any clinical surveillance system anywhere.

Zero.

We have no adverse event reporting system for AI mental health harm. We do not have a VAERS. We do not have an FDA MedWatch. We do not have a poison control center. We have newspaper articles.

That is the surveillance infrastructure for a technology that 5.4 million American kids are already using for mental health advice.

And then a Manhattan federal judge made it worse and better at the same time. Judge Rakoff ruled that conversations with consumer AI chatbots are not protected by attorney-client privilege. They are discoverable. They are evidence.

Every person pouring their darkest thoughts into a chatbot at 3AM is building a legal record they do not know exists.

The paper trail is forming. Not because anyone designed it. Because the courts, the researchers, and the reporters are doing the work that the platforms refused to do and the regulators have not figured out how to do yet.

The question is whether the paper trail gets long enough to matter before the next name gets added to it.

Because at the end of the day, we are all on the same side. AI assisted, but the human side.

AI Safety and Regulation

43 States Wrote 240 AI Bills in One Quarter. The White House Tried to Stop Them. It Did Not Work.

Manatt, Phelps and Phillips just published the most comprehensive AI health policy tracker in the country. The numbers are staggering. 43 states. Over 240 bills. In one quarter. That is nearly the entire output of 2025 compressed into three months.

The White House issued Executive Order 14365 in December directing the DOJ to form an AI Litigation Task Force to challenge what it called onerous state AI laws. The Commerce Department was ordered to publish an evaluation of those laws by March 11. That list never materialized.

What did materialize was a letter. Fifty Republican state lawmakers from 24 states wrote to the president telling him to stop blocking their AI legislation.

In Florida, the House Speaker refused to bring the governor's own AI Bill of Rights to the floor after White House outreach. The bill had passed the Senate 35-2. It never got read in the House. In Utah, the White House sent a letter opposing a state AI bill. The bill failed on March 6.

But everywhere else, the bills kept moving. Three template patterns are spreading state to state: the Illinois therapy ban model, the California chatbot disclosure model, and the Texas clinician disclosure model. States are copying each other's homework. The patchwork is becoming a quilt.

Source: Manatt Health AI Policy Tracker, week of April 10, 2026

Takeaway: 240 bills in 90 days is not a policy discussion. It is a policy avalanche. The only question is whether the laws can get specific enough to matter before the platforms outrun them.

For Legislators: Manatt's tracker is the single best resource for comparing your state's approach to every other state's approach. If your bill mirrors Illinois, California, or Texas, Manatt has already mapped the gaps. Use it.

Mental Health Policy

NPR Investigated What Happened After the Kaiser Strike. The Triage Team Went from Nine to Three.

In Issue 13, we covered the Kaiser Permanente strike. 2,400 therapists walked out for 24 hours. They said the walkout was about AI. NPR went back and asked what happened next.

The answer is worse than what they struck over.

At Kaiser's Walnut Creek facility, the triage team that once had nine licensed providers now has three. Screenings that used to be 10 to 15 minutes with a licensed clinician are now conducted by unlicensed operators following a script. Or an e-visit.

Ilana Marcucci-Morris is a licensed clinical social worker who worked triage at Kaiser's Oakland telepsychiatry hub since 2019. In May 2025, she was reassigned. Her colleagues say the downsizing is paving the way for AI to take their jobs.

Meanwhile, NPR found nearly 40 different AI documentation products now on the market for therapists. Blueprint summarizes sessions and updates electronic health records. Limbic operates across 63 percent of the UK National Health Service and serves clients in 13 US states.

Everyone draws the line between administrative AI and clinical AI. But when you replace a licensed triage clinician with a script reader, you have crossed from documentation into clinical gatekeeping. You just did it without calling it that.

Source: NPR, April 7, 2026

Takeaway: The strike was the headline. The staffing cuts are the story. When triage goes from licensed clinician to unlicensed script reader, the question is who decided that the first clinical touchpoint for someone in crisis could be handled by someone without a license.

For Legislators: Ask your state's largest health systems for triage staffing data, before and after AI tool adoption. If licensed positions were eliminated and replaced by unlicensed roles reading scripts, that is not efficiency. That is a scope of practice question.

Conversational AI Technology

The Adults Are Using AI for Therapy Too. The First National Survey Just Measured How Many.

In Issue 13, we covered the 5.4 million American youth using AI for mental health advice. That was the JAMA Network Open survey of 12-to-21-year-olds. Now JMIR Mental Health has published the adult counterpart.

Researchers surveyed 1,805 US adults aged 18 to 49 through a nationally representative panel in October 2025. They asked how many days per week people use generative AI or chatbots to deal with mental health concerns. They asked how much time they spend. They asked whether using AI changed how often they see a human professional.

The headline is the subset who are replacing human therapists entirely. Heavy AI users, defined as daily use or near-daily use for 30 minutes or more, are a measurable and identifiable group. They are not supplementing care. They are substituting for it.

Most respondents still prefer human professionals. But "most" is doing a lot of lifting in that sentence. The gap is closing. And unlike the youth survey, these are adults making informed choices with full agency. They are choosing the chatbot over the therapist. Nobody is tracking outcomes for those choices.

Source: JMIR Mental Health 2026;13:e88196

Takeaway: We spent a year focused on kids. The adults are doing it too. Nobody is writing bills to protect a 34-year-old from choosing a chatbot over a licensed therapist. And no one is tracking what happens when they do.

For Legislators: Youth chatbot bills are necessary. But the JMIR data shows adult AI mental health use is growing faster than the clinical evidence supporting it. Consider disclosure mandates for AI tools marketed for mental health support, regardless of user age.

Patient Safety and Ethics

71 Media Reports of AI Psychiatric Harm. 36 Unique Cases. Zero Verified Through Any Clinical Surveillance System.

A team at the University of Montreal just published a rapid scoping review in JMIR Mental Health. They collected every mass media report of psychiatric adverse events associated with generative AI chatbot interactions.

They found 71 unique news articles mapped to 36 unique cases, published between September 2025 and January 2026. Most coverage came from the United States. Headline tone was coded as alarmist in 29 articles. The outcomes described were severe: suicides, psychotic episodes, hospitalizations, self-harm.

Then they asked the question that matters. How many of these 36 cases were verified through any formal clinical surveillance system? Zero. Not the FDA. Not the CDC. Not any state health department. Not any adverse event database. Nothing.

The only surveillance system we have for AI-related mental health harm is journalism. Reporters are our MedWatch. Newsrooms are our poison control. They are not designed for it, not resourced for it, and not required to do it.

When a drug causes psychiatric harm, there is FAERS. When a medical device fails, there is MAUDE. When a vaccine produces a side effect, there is VAERS. When an AI chatbot contributes to a suicide, there is a newspaper article. Maybe.

Source: JMIR Mental Health 2026;13:e93040 (Chung, Bernier, Hudon, Universite de Montreal)

Takeaway: We cannot regulate what we do not measure. We cannot measure what we do not count. We are not counting.

For Legislators: Mandate adverse event reporting for AI-related psychiatric harm. Model it on FAERS or VAERS. Give it to the FDA, the FTC, or a new body. But build it. Because right now the only people documenting the harm are reporters, and that is not a system.

Digital Health Innovation

A Manhattan Federal Judge Just Ruled That Every AI Chat Log Is Discoverable. Your Therapy Chatbot Conversations Are Evidence.

In February, US District Judge Jed Rakoff issued a written opinion in United States v. Heppner. The ruling was specific: AI-generated materials shared with consumer chatbots are not protected by attorney-client privilege. They are not work product. They are discoverable.

The case specifically involved Anthropic's Claude. But the ruling applies to any consumer AI chatbot. ChatGPT. Replika. Character.AI. Woebot. Wysa. Every platform where people type their thoughts at 3AM.

Law firms are scrambling. Engagement letters are being rewritten. Clients are being warned that anything shared with consumer AI chatbots can be subject to legal discovery.

Now apply this to mental health. Millions of people are using AI chatbots to discuss depression, anxiety, suicidal thoughts, trauma. They are doing it because chatbots feel private. They feel like a journal that talks back. They are not private. They are a record. And Judge Rakoff just confirmed they are a record that courts can demand.

This cuts both ways. For the nine families suing OpenAI, discoverability is the weapon that forces the chat logs into evidence. For the millions of everyday users, it means every vulnerable conversation is one subpoena away from being read by a stranger.

Source: Manhattan Today, April 15, 2026

Takeaway: Privacy in AI therapy is an illusion. It always was. But now it is a documented, court-confirmed illusion. Your chatbot conversations are not confidential. They are evidence.

For Legislators: None of the 240 bills tracked by Manatt require chatbot platforms to warn users that their conversations may be subject to legal discovery. Add it. Every user deserves to know that the chatbot is not a therapist, not a friend, and not confidential.

What We Built

Casey: Voice-First AI-Native Mental Health EHR

Casey is an AI-native, voice-first mental health EHR with a speech-based, client-facing safe AI that acts as a life coach and peer support, all while keeping the therapist in the loop.

The data layer features the first HIPAA-compliant Neo4j Memory Graph, which builds persistent therapeutic context across months of daily sessions. Pre-FDA safety validation complete: 1.78 million stress test executions at 100% accuracy.

Campus-first launch with founding North Carolina state licensee. 50-state PC licensee model. $2.5M seed raise in progress.

Watch the Casey Demo →