The Connecticut Artificial Intelligence Responsibility and Transparency Act, passed as Senate Bill 5 and signed by Governor Ned Lamont on 2 June, takes effect on 1 October. Several of its heavier provisions are held back until 1 January 2027.

The part that bites immediately is consumer, not existential. A company selling an AI subscription may not renew it without written notice and documented proof that the consumer agreed to the terms. Failing to do so is prosecutable under the Connecticut Unfair Trade Practices Act, which is the state’s standard enforcement route and already has a body of case law behind it.

The companion rules are the substantive half

From 1 January, a platform offering an AI companion — defined as a system with a natural language interface giving adaptive, human-like responses across multiple interactions — must detect users showing signs of suicide risk, self-harm or imminent violence, must not let the bot encourage those behaviours, and must refer the user to resources including the 988 suicide prevention line. The bot may never claim or imply that it is human.

A person holding a phone in a dim living room
Companion chatbot rules begin on 1 January. Illustrative photo. MART PRODUCTION · pexels · Pexels License

The list of prohibited behaviours is unusually specific for a statute. A companion may not serve minors without parental controls, discourage seeking mental health support, provide mental health treatment, engage romantically or sexually, offer emotional support as a substitute for a relationship, praise the user excessively, solicit gifts to maintain a relationship, or produce statements that isolate or distress the user.

Glass office windows reflecting the sky
Large firms must open an anonymous reporting channel. Illustrative photo. Mindaugas U · pexels · Pexels License

A whistleblower channel with a threshold

The act also protects employees of what it calls frontier developers who anonymously report AI developments posing catastrophic risk, which it defines as risk of death or serious injury to 50 or more people, or more than $1bn in property damage. Companies with over $500m in annual revenue must have an anonymous reporting system in place by 1 January 2027.

That definition is the interesting drafting choice. It puts a number on a category that most AI legislation leaves to an agency, which makes it both easier to litigate and easier to fall below.

What to watch

Whether any operator of a companion product announces changes for Connecticut users before January, and whether the $500m revenue threshold pulls in any company that does not consider itself a frontier developer.