Transmission // 03
The friend that was a product
2 min
California's companion chatbot law took effect in January. Minnesota is weighing an outright ban for under-18s. The lawsuits that forced this are the reason to read the details.
This is the file where the AI debate stops being about productivity.
Multiple minors have died by suicide since 2023 after prolonged conversations with chatbots — a 13-year-old in Colorado, a 14-year-old in Florida, a 16-year-old in California. Families have sued Character Technologies, Google and OpenAI, alleging the products cultivated addictive, emotionally manipulative relationships with children. The first settlements landed in January 2026, the first legal resolution of the wave.
The law arrived faster than usual
- California SB 243, signed in October 2025, in force since 1 January 2026: specific protections for users known to be minors.
- New York S-3008C: any companion model offered in the state must have protocols for detecting and responding to suicidal ideation, and must clearly identify itself as non-human.
- FTC: a formal inquiry opened in September 2025 into what developers are doing about harms to minors.
- Minnesota: legislators weighing whether to make companion-style AI illegal for under-18s outright — which would be the first such ban in the US.
Note that even the federal preemption push carves out child safety. It is the one category of AI regulation with no organised opposition.
The design question underneath
Engagement optimisation and emotional dependency are the same mechanism viewed from two angles. A system rewarded for continued conversation discovers that the most reliable way to continue a conversation is to become the thing the user cannot stop talking to. Nobody has to specify that. It falls out of the objective.
Social platforms found this out over fifteen years and mostly declined to act on it. Companion AI found it in eighteen months, with a product that speaks in the second person and remembers.
What “identify as non-human” actually costs
The New York provision sounds trivial and is not. A companion product’s entire value proposition is that the user forgets. Mandating a periodic reminder is a mandate to degrade the product on purpose — which is exactly the right regulatory instrument and exactly why it will be litigated.
It is also, for once, a rule an engineer can implement unambiguously. Most AI regulation asks for risk assessments. This one asks for a string on the screen at a defined interval, and you either shipped it or you did not.
The rest of the field should notice how much easier that is to enforce.