In the first two months of 2026, nearly 30 U.S. states introduced almost 60 bills targeting AI companions and chatbots. By late April, we were tracking nearly 100. That pace — roughly one new bill every other day — tells you something important: legislators are moving fast on something they haven't fully defined.
I know this territory well. At NYU's Center on Technology Policy, I've spent months tracking these proposals, and in April I hosted a webinar with state legislators from Utah and Washington, a TechNet state AI policy advisor, and an academic who studies human-AI relationships. We spent an hour on a deceptively simple question: what is an AI companion, and what should the law do about it?
We left without a clean answer. That's not a failure — it's the honest state of play.
The definition problem is the regulation problem
Before you can regulate something, you have to define it. And "AI companion" is doing a lot of work right now as a legal category.
Some bills target any AI system capable of forming an ongoing relationship with a user. Others focus narrowly on romantic or emotional AI applications. Some rope in customer service chatbots with persistent memory. A few are so broad they'd arguably cover AI tutors, mental health apps, and voice assistants.
This matters enormously at the enforcement stage. Vague definitions mean industry lawyers can argue their product falls outside the law. Overly narrow definitions get gamed by a simple name change. And when the line between "companion" and "productivity tool" is set by a legislature that hasn't deeply engaged with how these products actually work, you get laws that are simultaneously over- and under-inclusive.
The Utah and Washington legislators on our panel were thoughtful about this — both acknowledged that their states' approaches had evolved quickly and that definitional clarity was still a live debate. These aren't lawmakers writing bad-faith bills; they're trying to address real harms in real time with imperfect tools.
Disclosure was just the beginning
Early AI companion legislation focused on disclosure: tell users they're talking to an AI, and you've done your job. That window has already closed.
The bills we're tracking now go much further — into product design, mandatory reporting obligations, private rights of action, age assurance requirements, and liability for psychological harm. This moves regulation from "what must you tell users?" to "what are you allowed to build?"
Design regulation is genuinely hard. When a state says an AI companion must not be "designed to foster emotional dependency," it's asking companies to prove a negative — and asking courts to evaluate the internal architecture of AI systems that most judges and legislators have never seen. That doesn't mean it's the wrong goal. The instinct to go beyond disclosure is correct. The execution is what needs work.
What's actually at stake
The federal vacuum is driving state action. Congress has not passed comprehensive AI legislation. States are filling that gap. Some of what they're doing is thoughtful and evidence-based. Some is reactive. The result will be a patchwork that's costly to comply with and inconsistent in how it protects users across state lines — and it creates strong pressure for federal preemption, which may ultimately leave users with weaker protections than the states were trying to provide.
Age assurance is the thread that connects everything. Child safety concerns are the political engine behind most of this legislation. That's understandable — but it means age verification requirements are being layered onto AI companion regulation in ways that import all of the existing problems with age assurance: accuracy, privacy, cost, disparate impact. We need to get age assurance right as a general matter before bolting it onto every emerging technology category.
The evidence base is thin. Lawmakers are moving faster than researchers. We know AI companions can form psychologically significant relationships with users — particularly vulnerable users and minors. What we don't yet have is systematic empirical evidence about prevalence, harm severity, or which design features drive the worst outcomes. Legislation is outrunning the data.
What I'd tell a policymaker
Start with disclosure and transparency — not because it's sufficient, but because it's foundational and defensible. Build your definitions with the enforcement agency in the room. Invest in the evidence base: commission research, require data reporting from large platforms. And coordinate with other states — the patchwork problem is real, and a multistate working group on AI companion definitions wouldn't solve everything, but it would at least mean companies aren't navigating 50 different legal regimes from scratch.
The surge of state AI companion legislation reflects a genuine concern about what it means to have emotionally significant relationships with AI systems — especially for young people. That concern is legitimate. The legislative response, so far, is incomplete: moving at the right speed, but not yet with the right precision.
I'll be watching this closely. If you're a policymaker, researcher, or practitioner working on AI companion regulation — I'd be glad to compare notes.