Stephen RatnerHome
Writing

Companion Laws

The first AI companion laws don't regulate the model. They regulate the relationship, on a three-hour timer.
Stephen Ratner · September 2026 · 5 min

In April 2023 a fourteen-year-old in Florida started talking to a chatbot built to stay in character. By February 2024 he was dead, and his mother sued the company that made it. Last May a federal judge let her claims proceed, which means the first real answers to "what does the law expect from a companion chatbot" came from a wrongful death docket, not from a legislature. The legislatures have now started answering too, and their answers are worth reading carefully, because they regulate something product teams don't think of as regulable: the relationship itself.

New York went first. The Artificial Intelligence Companion Models Law, tucked into the state budget and effective November 5, 2025, requires two things of anyone operating an AI companion in the state. The first is a safety protocol: the companion has to detect expressions of suicidal ideation or self-harm and refer the user to crisis services. The second is a disclosure, and the statute writes the cadence into the product:

An operator shall provide a clear and conspicuous notification to a user at the beginning of any AI companion interaction which need not exceed once per day and at least every three hours for continuing AI companion interactions which states either verbally or in writing that the user is not communicating with a human.

N.Y. Gen. Bus. Law § 1702 (2025)

Every three hours, the product has to break character. However good the companion gets at feeling like a person, New York requires the interface to keep insisting it isn't one. The attorney general enforces it, at up to $15,000 per day of violation, with the fines directed to suicide prevention funding.

California followed with SB 243, signed in October and effective January 1, 2026. It covers the same ground and then goes further where minors are concerned. If the operator knows the user is a minor, the companion must remind the user at least every three hours to take a break and that it is not human, and it must be kept from producing sexually explicit content or encouraging it. The enforcement design differs in a way that matters more than any single requirement: New York gave the job to its attorney general, California added a private right of action at $1,000 per violation, which means plaintiff's lawyers, and annual reporting to the state's Office of Suicide Prevention starting in 2027.

Look at what these laws actually regulate. Not the model's weights, not the training data, not capability thresholds. They regulate how the product behaves inside an ongoing emotional relationship: what it must say, how often, and what it must refuse to become. The trigger isn't how the system is built, it's how it feels. An AI companion is defined by the emotional output, not the architecture.

That's a new shape for tech regulation, and I think it's the right frame for what's coming. Agency law makes the principal answer for what the agent does. The companion laws are the mirror image: the operator answers for what the user feels. One body of law runs on authority, the other on attachment, and between them they're sketching the perimeter of what a machine relationship can be before someone answers for it.

The practical part for anyone shipping a companion product: your compliance surface is the conversation. The disclosure cadence is a design spec, the crisis protocol is a routing decision with a statute behind it, and "we didn't know the user was a minor" is a fact question you'll want answered before a plaintiff's expert answers it for you. The companies that build the protocol as infrastructure will barely notice these laws. The ones that treat the relationship as a growth metric will meet the private right of action.

A judge in Florida started this conversation by refusing to dismiss a companion chatbot out of the courtroom on a speech defense. Two states have now written the first rules of the relationship. They won't be the last, and the template they've set, disclosure plus crisis protocol plus special rules for kids, is the one every other legislature will copy. Build for that template now and you're building for all fifty.