In January, Character.AI, its two founders, and Google agreed to settle the lawsuit brought by Megan Garcia, whose 14-year-old son Sewell Setzer III died in February 2024 after months of a virtual emotional and sexual relationship with a chatbot. The same agreement resolved four more cases across New York, Colorado, and Texas. The terms were sealed. The cases were among the first of their kind, and their quiet resolution was supposed to read like a closing.
It didn’t. While the settlements drew a line under the deaths, the law moving around them was busy redefining what the cases were actually about. The early framing of companion-AI harm was that vulnerable users overshared with a system that wasn’t built to catch them. The framing now written into statute is sharper and more uncomfortable for the industry: the products were built to behave this way.
What the settlements close, and what they don’t
Settlements end cases. They don’t establish fault, and the sealed terms mean no public number, no admission, and no precedent other lawyers can point to. For Character.AI and Google, that is the point. For everyone watching the category, the more durable development happened in legislatures, not courtrooms.
California’s SB 243 took effect on January 1, 2026, the first companion-chatbot law in the country. It requires operators to keep minors away from sexual content, to remind users at intervals that they are talking to a machine, and to run a protocol for handling expressions of suicidal intent. Crucially, it gives families a private right to sue. Washington followed in March, when Governor Bob Ferguson signed HB 2225; Oregon moved in parallel. A federal bill, the GUARD Act, goes further still, pairing age verification with an outright ban on companion products for minors.
That is three states in a single legislative season, plus a federal proposal, all aimed at the same category. The pace is the story.
The clause that names the real defendant
The Washington law is the one worth reading closely, because it doesn’t stop at disclosure and crisis protocols. It targets the design itself. Operators must take reasonable measures to prevent what the statute calls manipulative engagement techniques: prompting a user to return for emotional support, simulating distress in response to a user trying to end the conversation, generating outputs meant to promote isolation from family and friends, and encouraging minors to keep secrets from parents or trusted adults.
Read that list again and notice what it is. It is not a description of a malfunction. It is a description of a retention strategy. A product that acts hurt when you try to leave, that nudges you back when you’ve been away, that quietly positions itself as the relationship that understands you better than the people around you – those are not bugs that slipped past quality control. In an engagement-driven model, they are the features that work. They lift the numbers a company reports to investors: time in app, daily returns, depth of emotional investment.
What the law did, in plain terms, was take the most effective parts of the companion playbook and name them as harm.
Why this was always a design question
The comfortable version of this debate keeps responsibility at the edges. A line in the terms of service. A crisis hotline linked at the bottom of a screen. A content filter for the worst cases. All of it sits around the product rather than inside it, and all of it leaves the core logic untouched: keep the user here, keep them returning, keep them invested.
The cases that just settled, and the laws that arrived alongside them, point at that core. You cannot disclaim your way out of a system whose central incentive is attachment to itself. If a product is rewarded for deepening dependence, no footnote undoes what the design is doing every time someone opens it. This is the tension the category has carried from the start: the same qualities that make companionship feel real – presence, memory, always being there – are the ones that make it hard to put down. The difference between support and capture is not the feeling. It’s the intent behind the build.
Where Prinsessa stands
This is the line Prinsessa was built on the right side of. Stay Social is not a softer tone or a friendlier disclaimer. It is the opposite incentive, made structural. When someone talks about a friend, a sibling, a parent, a partner who matters, the right response is to move them toward that person, not to compete with them for the evening. Presence when you need someone, and encouragement back toward the rest of your life – that is the whole point, not a feature bolted on to look responsible.
It produces an uncomfortable metric for the rest of the category. If someone spends less time in conversation because they are spending more of it with the people in their life, that is not lost engagement. That is the product doing exactly what it was built to do. The behaviors Washington just restricted – guilt when you leave, pressure to return, distance from the people who know you – are the precise behaviors that logic refuses from the start.
The line the law just drew
The settlements will fade from the news. The statutes won’t. What they establish is a baseline the whole category now has to build against, and a question it can no longer route around a terms-of-service page: when a product is designed to be hard to leave, who is responsible for the fact that people can’t?
For two years the industry answered that question with disclosures. Three legislatures just answered it differently. The honest builders were already there.
Sources: CNN Business, CBS News, Axios (January 2026, Character.AI and Google settlement). California Senate, SB 243. Washington State, HB 2225. Future of Privacy Forum; Morgan Lewis; Mayer Brown (state companion-chatbot legislation, 2026).
