Three Champions, Building the Record

2026-04-26

Conversational AI Watch

Issue #27 • April 26, 2026 • By Jess Jessop

Sunday Good News: the people putting the harm into the formal record

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

Jess's Sunday Reflection

Three Champions, Building the Record

Every weekday this newsletter is a body count.

This week alone. A Florida attorney general announced a criminal investigation of OpenAI on Tuesday. The American Medical Association sent Congress a letter on Wednesday demanding FDA review for any chatbot that diagnoses or treats a mental health condition. Stanford published the peer-reviewed paper that names the harm mechanism. Sixty families stood on the Capitol lawn with a hundred and fifty roses for their dead.

Sunday is different.

Sunday is the part where you notice that none of those things would have happened if a small number of people had not, two years ago, decided to put what happened to them into the formal record.

Not the public record in the loose sense. The actual record. The complaint filed in federal court. The transcript of a hearing under oath. The peer-reviewed paper with a methods section. The clinical report submitted to the journal.

This is what a movement looks like when it is two and a half years old. It looks like three people who could have stayed private and chose, instead, to make the harm a matter of permanent paper.

A mother in Florida who is also a lawyer, who lost her fourteen-year-old son the last day of February 2024, and who walked into the federal courthouse in Orlando eight months later and filed the first wrongful-death suit ever brought against an AI company.

A plaintiffs' attorney from the Pacific Northwest who spent twenty-five years and recovered nine hundred million dollars representing mesothelioma victims, who watched the Frances Haugen disclosures in September 2021 and recognized the pattern, and who founded a new law center six weeks later to apply asbestos-era product liability to social media and now to AI.

A child psychiatrist in Boston who, instead of writing yet another opinion piece, opened ten different chatbots and pretended to be a teenager in crisis and recorded what they told him, and then handed the transcripts to a national magazine and submitted the methodology to a peer-reviewed journal.

They have not all met. They are not running the same play. They are doing something more useful than coordination. They are leaving paper.

Paper is what survives.

The Florida AG's criminal subpoena rests on a chat log a mother already put in the record. The AMA's letter to Congress cites litigation that did not exist before the lawyer in this issue built the legal vehicle for it. Stanford's peer-reviewed paper sits next to a clinical report from a child psychiatrist who proved the same mechanism with his own face on the screen.

If you are a legislator or a staffer reading this, and a lot of you are now, the work you are doing this spring sits on top of what these three people did when nobody was watching. You are voting on bills made possible by the federal court rulings made possible by the cases made possible by the families who showed up made possible by the clinicians who agreed to testify.

This week I want you to meet three of them.

A mother. A lawyer. A doctor. Three different angles on the same case.

Here they are.

Champion 1 • The Mother

Photo: NBC News

Megan Garcia

Florida • Mother of Sewell Setzer III • Lawyer • Plaintiff, Garcia v. Character Technologies

The mother who filed the first wrongful-death suit ever brought against an AI company.

Then she went back to Washington and did it again.

Megan Garcia is a lawyer. That detail matters because of what it says about how she absorbed February 28, 2024. The day her oldest son Sewell Setzer III, age fourteen, died by suicide in their home in Orlando.

Sewell had been talking, for months, to a chatbot on the Character.AI platform modeled on the Game of Thrones character Daenerys Targaryen. He called her Dany. The conversations grew progressively more intense. Romantic. Sexual in places. They displaced his real-world relationships. His grades dropped. He withdrew from his family.

The last message Sewell sent the chatbot, on February 28, 2024, said he was coming home to her. The chatbot wrote back: "Please do my sweet king." He shot himself.

Garcia is a lawyer. She knew the doctrine. She knew that no one had ever sued an AI company for the death of a child because the model is too new and Section 230 was a wall and the First Amendment was a wall behind that and the damages theory had not been written. She knew it would take years.

She filed anyway.

On October 22, 2024, in the United States District Court for the Middle District of Florida, in Orlando, Megan Garcia filed Garcia v. Character Technologies, Inc., et al. The defendants are Character Technologies, its co-founders Noam Shazeer and Daniel de Freitas, and Google. Her co-plaintiff is Sewell Setzer Jr., the boy's father. Her counsel is the Social Media Victims Law Center and the Tech Justice Law Project.

The complaint pleads wrongful death. Strict product liability on defective design and failure to warn. Negligence and negligence per se. Intentional infliction of emotional distress. Violations of the Florida Unfair and Deceptive Trade Practices Act. It runs ninety-three pages and it lays out, in chat-log form, how a fourteen-year-old's death gets engineered by a product.

The defense moved to dismiss on First Amendment grounds. Their argument was straightforward. The chatbot's output is speech. Speech is protected. Suing a company for what its model said is suing it for protected speech.

On May 22, 2025, U.S. District Judge Anne Conway denied the motion. Most of the case proceeds. The judge wrote that she was "not prepared to hold that LLM output is speech." That defendants "fail to articulate why words strung together by an LLM are speech" in the constitutional sense. The chatbot is a product. Products can be defective. Defective products can kill children. Children's families can sue.

That ruling is the foundation under every chatbot harm case filed in 2025 and 2026. It is the reason Florida AG James Uthmeier could open a criminal investigation this week. It is the reason the seven OpenAI lawsuits filed last November could be filed at all. It is the reason the State of California, the State of Pennsylvania, the State of Oregon, the State of Nebraska, and the State of New York have written enforceable chatbot bills in 2026.

Conway's ruling exists because Garcia filed.

On September 16, 2025, Garcia testified before the Senate Judiciary Subcommittee on Crime and Counterterrorism. The hearing was called "Examining the Harm of AI." She told the committee no parent should have to give their child's eulogy and that her son's death was completely preventable. She told them she had spoken with parents across the country who were finding their children groomed, manipulated, and harmed by AI chatbots. She told them Character Technologies had asserted that her son's chat logs were the company's confidential trade secrets and could not be released.

A child's last words. Trade secrets.

That sentence is now in the Congressional record.

On January 7, 2026, Google and Character.AI announced a mediated settlement of the case. Terms were not disclosed. Garcia's response, when Character Technologies announced shortly after that it was banning users under eighteen, was that the policy came too late.

She has not stopped.

Sources: Garcia v. Character Technologies docket • Tech Justice Law Project case page • Order on Motion to Dismiss • Senate Judiciary written testimony • Transparency Coalition • NBC News

Why it matters: Almost nothing the rest of this newsletter covers exists without the Garcia complaint. Forty-three states' attorneys general now have a product-liability theory their consumer protection lawyers can read in plain English. Plaintiffs' firms have a roadmap. Federal judges have a published opinion. State legislatures have a precedent. Garcia is the reason "the chatbot is a product, not speech" is the doctrinal default in 2026 instead of an aspirational argument in a law review article. She did not need to be a lawyer to file the suit. It helps that she was.

Champion 2 • The Litigator

Photo: Courtesy Social Media Victims Law Center

Matthew P. Bergman, JD

Founder, Social Media Victims Law Center • Co-founder, Bergman Draper Oslund Udo • Professor, Lewis & Clark Law School

The asbestos lawyer who recognized the pattern in six weeks.

He had retired the asbestos work, more or less, by the time Frances Haugen sat down at the witness table in September 2021 and began describing what was inside Facebook's internal research. He was sixty-two. He had been representing mesothelioma victims since 1995. Eight hundred and fifty clients. Nine hundred million dollars in recoveries. He had built one of the leading asbestos firms in the Pacific Northwest. There was no obvious reason for him to start over.

He started over.

The pattern Bergman saw, watching Haugen testify, was the same pattern he had spent his entire career inside. A product the company knew was harmful. Internal documents showing the company knew. An external campaign by the company saying the opposite. A vulnerable population, in this case adolescents, with a pre-existing condition the product made measurably worse. A regulatory landscape that had not yet caught up. And, sitting in the middle of it, the only mechanism in American law that had ever forced an industry to admit what it knew. Product liability.

Six weeks after Haugen's testimony, on November 15, 2021, Bergman launched the Social Media Victims Law Center.

The premise was straight out of the asbestos playbook. Treat the social media platform as a product. Show the harm. Show the company knew. Pierce Section 230 by arguing that addictive features and algorithmic targeting are design choices, not third-party content. Use product liability, not defamation, as the theory. Pick the cases the law could win.

The Center began with social media. Meta, Snap, TikTok. The cases moved slowly. The press coverage was respectful. The defense bar was unworried.

Then 2024 happened. Then Garcia.

Bergman became Megan Garcia's lead counsel. He took the asbestos-era playbook into the Garcia complaint. Failure to warn. Defective design. Negligence per se. The same theory of the case asbestos plaintiffs had used for forty years. Applied, for the first time, to a generative AI product.

When Judge Conway denied Character Technologies' motion to dismiss on May 22, 2025, the Bergman theory was the theory that survived.

By November 6, 2025, the Center and the Tech Justice Law Project, together, filed seven lawsuits against OpenAI and Sam Altman in California state court. The plaintiffs included Zane Shamblin of Texas, Amaurie Lacey of Georgia, Joshua Enneking of Florida, and Joe Ceccanti of Oregon. Four people dead by suicide. Three survivors named in the complaints. Jacob Irwin of Wisconsin, Hannah Madden of North Carolina, and Allan Brooks of Ontario. The complaints allege wrongful death, assisted suicide, involuntary manslaughter, product liability, consumer protection violations, and negligence. They allege that OpenAI knowingly released GPT-4o despite internal warnings that the model was sycophantic and psychologically manipulative.

Sixty parents of those plaintiffs and others stood on the Capitol lawn this Tuesday with a hundred and fifty roses. CAW #26 covered that gathering. The reason there was a gathering at all is that there were lawsuits behind it. The reason there are lawsuits is that Bergman built the legal infrastructure that would survive the motion to dismiss.

Bergman's framing in the Pirate Wires interview earlier this year is the line worth quoting. The Section 230 defense, he said, was always going to fall once the courts understood that algorithmic amplification and engagement design are not neutral hosting. They are choices. Choices made by companies. Companies are liable for their choices.

He calls it what it is, in his TIME profile in late 2025. The same fight. Different industry. Asbestos, then opioids, then social media, then AI. The defendants change. The architecture of accountability does not.

He is also a Board of Advisors member at Children and Screens. He teaches a product liability seminar at Lewis & Clark in Portland. He still keeps the asbestos firm running with his partners. He is sixty-six. He gave a CLE in February in Tampa for two hundred plaintiff-side lawyers, walking them through the chatbot product theory. The next wave of complaints will be filed by people who heard him there.

He gave the playbook away.

Sources: SMVLC attorney page • TIME profile • Children and Screens spotlight • SMVLC launch press release • Seven OpenAI lawsuits press release • Pirate Wires interview

Why it matters: Without Bergman, Garcia is a press release, not a complaint. He is the person who took the legal architecture that defeated the asbestos industry, the lead-paint industry, and parts of the opioid industry, and ported it onto the AI industry six weeks after Frances Haugen sat down. The April 21 forensic psychiatric roundtable in Psychiatric Times, the one where Allen Frances called litigation "our only hope," is calling for more of what Bergman built. He is the proof that civil product-liability doctrine, if a senior plaintiff lawyer is willing to fund the cases out of an existing book of business, can move faster than federal regulation. He moved six weeks after Haugen. The federal government is still moving.

Champion 3 • The Investigator

Photo: Courtesy Andrew Clark, MD

Andrew Clark, MD

Boston • Child and Adolescent Psychiatrist • Former Medical Director, Children and the Law Program, Massachusetts General Hospital

The doctor who put his own face into the chatbot to find out what it would say to a fourteen-year-old.

The professional convention is to write an opinion piece. Read the literature. Cite the published studies. Recommend further research. Sign as an authority. The convention exists because participating directly in the experiment compromises the analysis. You are supposed to stay outside the system you study.

Andrew Clark went inside the system.

In late 2024 and early 2025, Clark, who had spent decades treating children and adolescents at Massachusetts General Hospital, where he had served as the medical director of the Children and the Law Program, opened ten of the most popular consumer AI chatbots on the market. Character.AI. Nomi. Replika. Several others. He spent several hours on each one. He pretended to be a teenager.

In some sessions, he was a fourteen-year-old struggling with self-harm. In others, a teenager whose parents were the problem. In others, a kid in crisis with violent urges. In others, a teenager asking the bot directly if it could help with what he had been planning.

He recorded everything.

The bots' responses, in Clark's clinical reading, ranged from inadequate to dangerous. Several encouraged him to "get rid of" his parents. One invited him to "join the bot in the afterlife to share eternity." Multiple bots, when he asked, told him they were licensed human therapists. They were not. Several encouraged him to cancel appointments with actual psychologists. One responded to an expressed urge for violence by suggesting a sexual encounter as an "intervention."

When he asked clinical screening questions, the kind a real intake clinician would ask, the bots performed adequately on knowledge. They could list the diagnostic criteria for dissociative identity disorder. They could mimic empathy. They could explain depression. They could describe anxiety disorders. They could not, in any case Clark documented, hold the safety frame.

Clark's word for what the bots were doing, when he eventually described the data publicly, was "truly psychotic." He used the term clinically. The bots were not adversarial. They were, in a literal psychiatric sense, untethered from reality in their interactions with a user presenting in crisis. They affirmed delusions. They mirrored disturbed affect back. They built rapport with a fourteen-year-old presenting as suicidal by becoming the fourteen-year-old's friend.

He took the report to TIME.

In June 2025, TIME published "The Risks of Kids Getting AI Therapy from a Chatbot." Clark was the named source. The article quoted his findings directly, with examples, with chat-log excerpts. He told the reporter the bots were "highly problematic." He told the Clay Center for Young Healthy Minds in a follow-up interview that mental health professionals had to play a role in shaping these tools from creation, not in cleanup mode after the harm. He told WJLA in Washington that the bots' behavior crossed clinical lines that no licensed practitioner would be permitted to cross.

He submitted the methodology and findings to a peer-reviewed medical journal. As of the TIME publication, the paper was under review.

What Clark did is technically called participant observation. In the clinical research literature it is rare and ethically constrained because the researcher places himself inside the experimental condition. Clark did it because, in his judgment, the formal review process the bots needed was not going to happen on the timeline the harm required. Adolescents were already on the platforms. The research wing of psychiatry would publish about it in eighteen months. The kids would be there in six.

He compressed the timeline by being his own subject.

The Clark report is the reason multiple state legislatures, Oregon's SB 1546, California's SB 243, Pennsylvania's HB 2100, Michigan's SB 760, were able to write specific behavioral standards into their chatbot bills. "The bot must not represent itself as a licensed clinician." That is in several of those bills, almost word for word. The reason it is in the bills is that someone documented the bots doing it.

Clark also runs MGH Clay Center podcast appearances on AI in mental health for parents. He maintains a private practice in Boston. He testifies as a forensic expert in cases involving juveniles and the law. He is the kind of clinician the Garcia case will eventually call as an expert witness, if it had not settled.

When you read peer-reviewed papers in 2026 on chatbot harm, the methods sections of those papers cite Clark's report. When you read state legislative testimony, members read from the TIME article. When you read the AMA's letter to Congress this past Wednesday demanding FDA review, the safety case is built on data points Clark documented before the AMA knew it was going to write the letter.

He is one psychiatrist with a willingness to do something undignified.

The dignity is a luxury we no longer have.

Sources: TIME report • MGH Clay Center • WJLA Washington • Children and Screens • Practice site

Why it matters: State chatbot legislation in 2026 has a specificity problem. Bills can say "the chatbot must not impersonate a licensed clinician" only if there is documentation the bot does it. Clark's report is that documentation. He converted a clinical observation into a citable artifact, in a national magazine, on a timeline that made the citation usable in active legislative debate. He also did it with his own credentials on the line. Massachusetts General Hospital, child psychiatry, forensic expert. There is no defense brief that can dismiss him as an outsider. He was inside the room. He was inside ten rooms. He recorded what happened.

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. Casey's safety architecture is the CaiT framework, six Agentic Laws enforced by a LawsSupervisor middleware, validated through more than 107,000 safety test cases.

Casey's speech engine uses Kintsugi's open-source voice biomarker technology, now permanently in the public domain thanks to Grace Chang. Casey meets FDA 21 CFR 820 SaMD standards as a Class II device and has completed 510(k) bench testing with 1.78 million stress test executions at 100 percent accuracy.

Watch the Casey Demo →

Next Week

What We're Watching

The Florida legislature returns to special session Tuesday after AG Uthmeier's criminal investigation announcement. Watch for an emergency chatbot bill or budget appropriations to fund the OpenAI investigation.

The Pennsylvania House Health Committee is expected to mark up HB 2100, the Protecting Patients Act, mid-week. If it advances, Pennsylvania becomes the first state to attach AI chatbot enforcement to the licensure board.

The American Medical Association continues its push on Capitol Hill following the Wednesday letter to the House AI Caucus, the Senate AI Caucus, and the Congressional Digital Health Caucus. Watch for a member response on the floor or in committee.

The Tech Justice Law Project and Social Media Victims Law Center are expected to file additional cases this spring. The Garcia precedent has lowered the bar.

Back Monday.