A new Connecticut state law regulating artificial intelligence and AI chatbots is going into effect on Thursday.
The law is one of more than 90 laws going into effect this week, concerning topics from immigration to traffic safety to gun laws.
The law prohibits AI companies with subscription services — such as ChatGPT or Google AI, for example — from renewing subscriptions without written notice and proof that the consumer has agreed to the terms and conditions of the subscription.
Companies who do not adhere to these requirements will be prosecuted under the Connecticut Unfair Trade Practices Act, which prevents people and companies from engaging in deceptive and unfair acts of business.
The law also enacts protections for whistleblowers from AI companies from being fired or otherwise retaliated against for reporting company activities that may endanger public health or safety.
The law is specifically referring to “frontier developer” employees, defined by the law as anyone who is training a machine-based system that can generate outputs and be trained in a variety of distinctive tasks.
Under the new law, frontier development employees must be allowed to anonymously report any AI development or outputs that might pose a “catastrophic risk” to the public as in “materially contribute to the death of, or serious injury to, more than 50 people, or more than $1 billion in damage to or loss of tangible or intangible property, from any single incident,” according to the General Assembly’s law summary.
An example of an AI development or output contributing to catastrophic damage would be if the bot provided “expert-level assistance in creating or releasing” some sort of biological or nuclear weapon or engaging in conduct “with no meaningful human oversight, intervention or supervision” that would be considered a cyberattack or extreme crime like murder or assault if it had been done by a human.
Large frontier development companies — those that saw more than $500 million in gross revenue last year — must have these anonymous reporting systems running by Jan. 1, 2027, according to the law.
Another part of this law establishes the Connecticut Artificial Intelligence Responsibility and Transparency Act, otherwise known as the CART Act. This statute requires that companies providing “AI companions” — defined by the law as “AI format with a natural language interface that gives adaptive, human-like responses to user inputs and can sustain a relationship across multiple interactions,” — have protocols to detect when chatbots users are “clearly showing a risk of suicide, self-harm, or imminent physical violence.” From there, the companions must have built-in prevention measures to ensure the bot doesn’t encourage the behavior in any way. This specific part of the law goes into effect on Jan. 1, 2027.
Part of the companion’s response, the law says, must refer users to mental health evaluation and treatment resources like the 9-8-8 Suicide Prevention Hotline.
The law also instructs AI companion operators to ensure that the bot never claims or implies it is human, even when directly asked by users. Companies and companion operators must clearly state that the bot is not human.
Furthermore, the CART Act prevents AI companies from providing AI companions to minors, unless there are means for parental control.
The bots are also not allowed to encourage self-harm or suicide, discourage users from searching for mental health support or offer mental health support itself, encourage harm to others or engage in a romantic or sexual relationship with the user.
The law specifically prohibits operators from allowing the bots to “engage” users in a variety of ways, including telling the user to use the bot as a means of emotional support or for a romantic bond, “excessively praising the user” or soliciting gifts or purchase to maintain the “relationship” with the bot. The bots also shouldn’t be able to output statements that might isolate or cause emotional distress to the user, the law says.