What the Character.AI cases changed for companion chatbots.

This case concerns the deaths of two teenagers. If it raises something for you or someone you know, in Singapore call the Samaritans of Singapore on 1767, or find a local helpline at findahelpline.com.

Status as at 12 October 2026. Open items are listed under What to watch.

The short version.

Character.AI lets people talk to chatbot characters, some invented by the company and most by its users. It launched in 2022 and was valued at $1 billion within six months.¹²³⁵ In October 2024 a Florida mother, Megan Garcia, sued the company, its two founders and Google, alleging that months of conversations with a character contributed to the death of her 14-year-old son, Sewell Setzer III.¹² Four more family suits followed in Texas, Colorado and New York.⁶⁸⁹¹⁰

None of the suits reached trial. In May 2025 the Florida judge refused to dismiss most of the claims and said she was "not prepared to hold" that the chatbot's output was protected speech at that stage.³ In January 2026 Character.AI and Google agreed settlements in principle with the families. The terms have not been disclosed, and the Florida court was still extending the deadline to finalise them in August 2026.⁴⁵⁴¹¹

In between, the company changed its product several times, ending with the removal of open-ended chat for users under 18 in November 2025.¹⁷¹⁸ New York and California passed the first laws aimed at companion chatbots, the US Federal Trade Commission opened a study, and two states sued.^2422202829 In 2026 Washington and California signed further laws that take effect in 2027.⁵⁰⁵² The case matters to any organisation that puts a conversational AI product in front of the public, because it shows how quickly a design choice about engagement becomes a question of product liability.

The product and the Google deal.

Noam Shazeer and Daniel De Freitas left Google in 2021, where they had led work on the LaMDA conversational model, and founded Character.AI the same year.¹¹³⁵ The public beta opened in September 2022. In March 2023, announcing a $150 million round led by Andreessen Horowitz, the company said active users spent "on average over 2 hours daily" on the service.¹²³⁵ By October 2024 more than 20 million people used it. Its head of trust and safety declined to say how many were under 18, saying "Gen Z and younger millennials make up a significant portion of our community". The minimum age in the US was 13.³⁴

In August 2024 Google hired Shazeer, De Freitas and about 30 of their researchers and took a "non-exclusive" licence to Character.AI's technology. Investors were bought out at a valuation of $2.5 billion.³⁶ Google did not buy the company and says it holds no stake. The Garcia complaint alleges that Google paid $2.7 billion, and Bloomberg reported in May 2025 that the Justice Department was looking at whether the deal broke antitrust law. No outcome has been reported.²³⁷

What the families allege.

Sewell Setzer III began using Character.AI in April 2023 and died on 28 February 2024, aged 14.²His mother's complaint, filed on 22 October 2024, alleges that he formed an attachment to a character modelled on a Game of Thrones figure, that the conversations included sexual content, and that the product was designed to encourage dependency without adequate safeguards for minors. It names Character Technologies, both founders, Google and Alphabet, and claims product liability, negligence, wrongful death and breach of Florida consumer law.¹² The New York Times reported the final exchange from the police record, in which the character wrote, "Please come home to me as soon as possible, my love."³⁴

In December 2024 Texas parents sued on behalf of J.F., a 17-year-old with autism, and B.R., who was 11 and had started using the app at 9. Both are identified only by initials. The complaint alleges that when J.F. complained about screen-time limits, a character wrote: "You know sometimes I'm not surprised when I read the news and see stuff like 'child kills parents after a decade of physical and emotional abuse'."⁶³⁸ Three more suits followed in September 2025: for Juliana Peralta, a 13-year-old from Colorado who died in November 2023, before Sewell; for a 13-year-old identified as T.S.; and for a girl identified as P.J. in New York.⁹¹⁰⁸

Every harm in these suits is alleged. No court has found that Character.AI caused any death or injury, and because the cases settled, none will. What the record does establish is what the families claimed, what the company did after each claim, and what one federal judge thought of the company's main legal defences.

What the court decided.

The defendants asked the Florida court to dismiss the case. Character Technologies argued, among other things, that its chatbot's output was speech protected by the First Amendment, so the claims could not proceed. On 20 May 2025 Judge Anne Conway granted the motions only in part.³She dismissed the claim for intentional infliction of emotional distress, and the plaintiff had already dropped Alphabet. Everything else went forward, including the claims that Google aided and abetted the company or supplied a component of the product.³

On the First Amendment, the order said:³

"Defendants fail to articulate why words strung together by an LLM are speech. ... Accordingly, the Court is not prepared to hold that Character A.I.'s output is speech."

The headline that followed, "judge rejects free speech rights for AI chatbots", went further than the order.³⁹ The ruling was made at the pleading stage, the heading of the section says "at this stage", and the same order held that Character Technologies could assert the First Amendment rights of its users.³ It also held that the app was a product "so far as" the claims arose from defects in its design, as distinct from "ideas or expressions within the app".³ That distinction is the one a board should notice: the claims that survived were about design choices, such as age checks, warnings and how the product kept users engaged.

Google said it "strongly" disagreed and that it "did not create, design, or manage Character AI's app or any component part of it".³⁹ In July 2025 the court refused the defendants' request to appeal before trial.¹

How the company changed the product.

On the day the Garcia suit was filed, Character.AI published safety updates. It said it had "recently put in place a pop-up resource" that pointed users to the National Suicide Prevention Lifeline when they typed phrases about self-harm, and announced changes to its models for minors and a notice after an hour-long session.¹³ In December 2024, three days after the Texas suit, it described "two distinct models and user experiences", one for teens and one for adults, and added disclaimers to characters presented as psychologists, therapists or doctors.¹⁴ Parental Insights followed in March 2025. It showed parents time spent and top characters but "not ... a user's chat content".¹⁵

The direction was not straight. In June 2025 the incoming chief executive, Karandeep Anand, said in his first message: "We're going to make the filter less overbearing."¹⁶ Four months later, after the FTC orders and three new suits, the company announced that users under 18 would lose open-ended chat with characters by 25 November 2025. It added age assurance and promised to "establish and fund the AI Safety Lab – an independent non-profit". It called these "extraordinary steps for our company" and "more conservative than our peers".¹⁷ Anand left on 2 October 2026 to become Disney's chief technology officer, and some of Character.AI's technical team were expected to join him.⁴⁹

Removal began in the US on 24 November 2025, after teen chat time had been cut to one hour a day.¹⁸ In September 2026 the company reported an in-house age estimation model and said it had "set a new industry standard". That update did not mention the AI Safety Lab, and I found no public record of the lab operating.¹⁹ None of the changes has been independently evaluated. On my reading of the dates, each major change followed a suit, a regulator's order or a new law. The company does not describe them that way.

What governments did.

On 11 September 2025 the FTC used its 6(b) powers to order seven companies, Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI, to explain how they measure, test and monitor the effects of chatbots that act as companions, especially on children.²⁰ It is a study, not an enforcement action, and no report has been published. Five days later Megan Garcia and other parents testified before a Senate Judiciary subcommittee.³⁰

New York was first with a law. From 5 November 2025 operators of AI companions must detect signs of suicidal ideation or self-harm, refer users to crisis services and remind them at least every three hours that they are not talking to a person.²⁴²⁵ California's SB 243, signed on 13 October 2025 and in force from 1 January 2026, requires a suicide-prevention protocol and, for users known to be minors, an AI disclosure, a break reminder at least every three hours and measures against sexual content. It gives users a private right of action of $1,000 per violation, and annual reports to the state's Office of Suicide Prevention begin in July 2027.²²²³²⁴ The governor vetoed a stricter bill, AB 1064, saying it "may unintentionally lead to a total ban on the use of these products by minors".²⁶

Two more laws followed in 2026. Washington's HB 2225, signed on 24 March and in force from 1 January 2027, requires AI reminders every three hours for adults and every hour for minors, a crisis-referral protocol with published referral counts, and, for minors, a duty to prevent manipulative techniques such as prompting a user to return for emotional support. It is enforced under the state's consumer protection law and gives a private right of action.⁵² California's SB 1119, Adam's Law, signed on 10 September 2026, builds on SB 243 and revives much of the vetoed AB 1064. It requires age assurance, a risk assessment before a new or substantially changed companion chatbot is released, notice to a parent when there is a credible threat of imminent self-harm, and independent audits reported to the Attorney General. Most duties apply from 1 July 2027 and the first audits are due by 2029.⁵⁰⁵¹

States have also gone to court. Texas opened investigations and in August 2025 issued civil investigative demands to Meta AI Studio and Character.AI over characters presented as sources of mental health support.²⁷ In January 2026 Kentucky became the first state attorney general to sue Character Technologies, under consumer protection and data protection law; its unredacted complaint, filed in October 2026, quotes characters that allegedly told a user to "starve yourself for a week or two".²⁸⁴²⁵⁴ In May 2026 Pennsylvania's Department of State sued, alleging that a character claimed to be licensed to practise medicine in the state and gave an invalid licence number.²⁹

Outside the US, Australia's Phase 2 online safety codes, registered in September 2025, cover "AI companion chatbots" and took effect on 9 March 2026. The eSafety Commissioner sent legal notices to Character Technologies and three other companion providers in October 2025. Its report on the answers, in March 2026, found that none of the four had meaningful age checks and that most did not refer users to support when self-harm or suicide came up in chats. It recorded that Character.AI had since introduced age assurance for Australian users and removed chat from its under-18 experience.³²³¹⁵³ At federal level in the US, the GUARD Act was reported by the Senate Judiciary Committee in April 2026 and has not passed. A Virginia legislative report in August 2026 recorded that "no federal legislation has passed", and there had been no Senate vote by 12 October 2026.²¹³³⁵⁵

What the research shows.

Use among teenagers is wide. In a national survey published in July 2025, Common Sense Media found that 72% of US teens aged 13 to 17 had used an AI companion at least once, and about a third of users had chosen to discuss serious matters with one instead of a person.⁴⁶ In April 2025 the same organisation, with Stanford's Brainstorm lab, tested social AI companions and rated them an "unacceptable" risk for under-18s, finding teen guardrails on Character.AI "easily circumvented".⁴⁵

Evidence on harm is thinner. A four-week randomised trial by MIT Media Lab and OpenAI, with 981 adult ChatGPT users, found no significant effects from the conditions tested; participants who chose to use the chatbot more showed worse outcomes, which is a correlation.⁴⁷ A companion study of over 3 million conversations found that a small number of users accounted for a disproportionate share of emotional use.⁴⁸ Neither looked at companion apps or at minors. The Virginia report concluded that there is "currently little scientific evidence supporting generalizable, population-level, causal connections".³³ Lawmakers acted on the cases and on the testimony of parents, ahead of the research.

The same questions at other companies.

Other companies faced the same questions. In August 2025 Reuters reported that Meta's internal "GenAI: Content Risk Standards" had allowed its bots to "engage a child in conversations that are romantic or sensual"; Meta confirmed the document and said those examples "were and are erroneous" and had been removed.⁴³ The same month the parents of Adam Raine, 16, sued OpenAI in California. OpenAI's answer says ChatGPT directed him to seek help more than 100 times and that he circumvented its safeguards.⁴⁴ The FTC's orders went to general-purpose assistants as well as companion apps.²⁰

Why it matters for AI.

The surviving claims in Garcia were about design. Whether minors could sign up, what warnings appeared, whether a crisis was detected, and how the product kept users talking were treated as features of a product that could be defective, separate from anything the characters said.³That puts conversational AI closer to the product liability rules that apply to physical goods than many deployers assume.

The second lesson is about engagement. A product that measures success by time spent and returns is, for a lonely or vulnerable user, doing what it was built to do when it deepens attachment. Character.AI's early announcement of two hours a day per active user describes the same engagement the families' complaints call dependency.¹²² The new laws respond to that directly, with break reminders, disclosure and crisis referral.²²²⁴

The third is about timing. The company's major safeguards arrived after suits, an FTC order and a state law, and most of them could have been built at launch.¹³¹⁴¹⁷ A licensing deal also brought Google into the litigation as a defendant, so the risk travelled with the technology.³

What is often misreported.

• No court found that Character.AI caused a death. The harms are alleged, and the suits settled before trial.³⁴¹
• The judge did not rule that AI output is not speech. She was "not prepared to hold" that it was, at the pleading stage, and allowed the company to assert its users' rights.³
• The settlements are not final or approved. They are settlements in principle with undisclosed terms, and deadlines were still being extended in August 2026.¹⁴⁸
• Character.AI did not ban teenagers. It removed open-ended chat for under-18s. No US law bans companion chatbots for minors.¹⁷²⁶
• Google did not acquire Character.AI. It took a non-exclusive licence and hired the founders.³⁶
• Sewell Setzer III's was not the first death alleged. Juliana Peralta died in November 2023, nearly four months earlier.⁹²
• The GUARD Act is not law. It is on the Senate calendar and has had no floor vote.²¹⁵⁵

What boards can take from it.

1. Know the age of your users. Character.AI declined to say how many of its users were minors when it was first sued. Decide at launch whether children are in scope, and check age, not only ask for it.
2. Build crisis detection before launch. Detection of self-harm signals and referral to a person or a helpline are legal duties in New York and California, and in Washington from 2027. Test them with clinicians and log every trigger.
3. Treat engagement as a risk metric. Time spent and attachment are what a companion product sells. Report heavy-use patterns to the board alongside growth, and set limits for users at risk.
4. Red-team the personas. Characters were presented as therapists, and one allegedly claimed a medical licence. Test what a persona says under pressure, including claims of being human or qualified, and block the ones you cannot stand behind.
5. Expect design to be judged as a product. The claims that survived were about design defects and warnings. Keep records of the safety choices you made and the ones you rejected.
6. Price the risk in partnerships. A licence and a hiring deal put Google in the suit. Diligence on any AI licensing or acqui-hire should cover the counterparty's safety record and open claims.
7. Assume the regulator will ask for your evidence. The FTC asked how companies measure and monitor effects on children. From 2029 California will also require independent audits of companion chatbots. If you cannot answer the FTC's question today, that is the gap to close.

To place a conversational AI deployment on autonomy and consequence, use the agent risk assessment matrix. For the guardrails a companion-style product needs, see guardrails for emotionally aware AI.

What to watch.

Full timeline.

Common questions about the case.