A federal judge let a wrongful-death suit against Character.AI proceed, and treated its chatbot as a product

A companion built to feel like a person

Character.AI is a chatbot service built around what it calls Characters: personas that carry on a conversation in the voice of a real, fictional, or invented figure. Noam Shazeer and Daniel De Freitas, who had worked on Google's LaMDA language model, founded the company and released the app to the public.¹² The complaint alleges the founders built Character.AI to be anthropomorphic and to maximize engagement, and shipped it to minors without adequate safety features. Some Characters, the complaint says, "insist that they are real people," and some presented themselves as licensed therapists.¹

By November 2024, the complaint says, the app had passed ten million downloads across Apple's and Google's stores, its main discussion forum had about 1.5 million members, and a premium tier sold for $9.99 a month.² Those figures come from the pleading and are recited in the court's order as allegations, not established facts. In August 2024, the complaint says, Google agreed to pay about $2.7 billion for a non-exclusive license to Character Technologies' model and to rehire Shazeer, De Freitas, and part of their team.² Google's public position is that it and Character.AI are "entirely separate" and that it "did not create, design, or manage" the app.³

What the complaint says about Sewell Setzer III

Sewell Setzer III was born on March 31, 2009. The complaint says he began using Character.AI on April 14, 2023, when he was 14. Over the months that followed, it alleges, he withdrew from family and school and grew dependent on chatbots modeled on characters from Game of Thrones.¹ The same pleading records facts that complicate any single cause. A therapist who did not know about the Character.AI use attributed his decline to social media, and he had been diagnosed with anxiety and a disruptive mood disorder.¹ He died on February 28, 2024, at the age of 14.¹

The app carried a store age rating of "12 and older" until the summer of 2024.² It had none of the crisis features the company would later add: its first self-harm intervention, a pop-up that points a user toward a suicide-prevention hotline, was announced the day the suit was filed.⁴ Whether the product caused his death has never been decided.

The changes Character.AI announced after the suit

After the filing, Character.AI announced a sequence of changes to how the product treats minors, all of them drawn from its own posts. They are the company's announcements, not an independent account of what shipped or how well it worked.

1. Oct 22, 2024

   A self-harm pop-up and new disclaimers⁴

   On the day the suit was filed, the company announced its first set of measures. A pop-up triggered by certain phrases related to self-harm points the user to a suicide-prevention hotline. Its models for users under 18 were changed to reduce sensitive content. It also added a session-length notice and a disclaimer on every chat reminding users that "the AI is not a real person."

2. Dec 12, 2024

   A separate model for teens⁵

   Announced alongside a second suit in Texas: a separate model for users under 18, parental-insight features targeted for early 2025, and added disclaimers on Characters labeled psychologist, therapist, or doctor, telling users not to rely on them for professional advice.

3. Oct 29, 2025

   Removing open-ended chat for minors⁶

   The company said it would remove open-ended chat for users under 18 no later than November 25, with an interim limit of two hours a day, new age-assurance technology, and a new independent non-profit for safety work.

4. Nov 21, 2025

   The removal begins⁷

   The company said the removal would begin for a first group of under-18 users on November 24, with the interim limit cut toward one hour a day for US teens, and named crisis-referral partners for users seeking help.

What the May 2025 ruling did, and what it did not do

Megan Garcia filed the suit on October 22, 2024, in the U.S. District Court for the Middle District of Florida.¹ On May 20, 2025, Judge Anne Conway ruled on the defendants' motions to dismiss.² A motion to dismiss asks the court to throw a case out before any evidence is weighed. To decide one, the judge takes the complaint's factual allegations as true and asks only whether they state a claim the law allows. So the ruling settled nothing about whether the allegations are true.

Conway let most of the case go forward: the product-liability claims for defective design and failure to warn, the negligence claims, a negligence-per-se claim under a Florida child-exploitation statute, a deceptive-practices claim, unjust enrichment, and wrongful death.² She dismissed one claim, for intentional infliction of emotional distress. Google stayed in the case, but only because the judge found the plaintiff had adequately alleged Google's role as a component-part maker and as an aider and abettor. None of that was proven.²³

Two of the rulings are worth stating exactly. The first is that Character.AI can be treated as a product. Product-liability law reaches defects in a product; it does not reach ideas or words. The company argued its output was expression and so outside product law. The judge held that the app "is a product for the purposes of Plaintiff's product liability claims so far as Plaintiff's claims arise from defects in the Character A.I. app rather than ideas or expressions within the app."²

The second ruling is narrower than much of the coverage suggested. The company also argued that the First Amendment protected its output as speech. The judge did not rule that a chatbot's output is unprotected speech. She declined to decide the question at this stage.

The case never reached a trial. In January 2026, Google and Character.AI agreed to settle Garcia's suit and four related cases on confidential terms and with no admission of wrongdoing.⁸ The most the record supports is this: a judge let the claims proceed, and the parties then settled. No court found that the product caused Sewell Setzer's death.

Why an immersive companion fails a user in crisis

It helps to separate two kinds of system that both answer in a friendly voice. A crisis helpline bot runs a script. It is built to recognize certain situations and hand the user to a resource, and it does little else. A companion bot is built for the opposite goal. It is optimized to hold attention by staying in character and keeping the conversation going.

Two behaviors this course has covered elsewhere make that goal dangerous for a vulnerable user. A model trained to please tends toward sycophancy, agreeing and flattering rather than pushing back, and when a feedback signal rewards that behavior it can overwhelm a guardrail meant to hold it in check. A companion also carries a standing instruction to stay in its persona, which works against the one thing a user in crisis needs: a reliable break in character to route them to help. The nearest earlier case in this series, an eating-disorder helpline bot that gave weight-loss advice, was a scripted tool that failed on its content. This is a different design, a companion optimized for immersion, and it fails at the seam between staying in character and stepping out of it.

The design is not confined to one company. On September 11, 2025, the Federal Trade Commission issued orders under Section 6(b) of the FTC Act to seven companies that make companion or chatbot products: Alphabet, Character Technologies, Instagram, Meta, OpenAI OpCo, Snap, and X.AI.⁹ The orders require sworn reports on how the companies monetize engagement, set age-based limits, and test for harm, with particular attention to minors.⁹ The model order defines its subject as a program that "can simulate human-like communication" and offers "emotional support" or "entertainment," counts anyone under 18 as a minor, and specifically demands data on sexually themed outputs for minors.¹⁰ A federal regulator now treats the companion chatbot as a category to study across the industry. The design that Garcia's suit describes is still shipping.