Connecticut's AI Companion Law (SB 5): What Changes and When

Two dates, two different obligations

SB 5 is a 74-page omnibus. Most of it deals with frontier-model whistleblowers, automated employment decisions, synthetic media and state workforce programmes. Two sections matter to anyone who pays for or uses a companion app, and they start on different dates. Some coverage collapses them into a single "1 October" effective date. The Act's own section headings do not.

Already in force: the AI subscription rule

Section 1 is the part that has applied since 1 October. A subscription-based provider may not enter into or renew a subscription, or collect a fee, unless it has given the consumer written notice of the key terms and the consumer has given written notice accepting them. For a first subscription, the notice must cover any quantitative or qualitative limits the provider may impose, including limits triggered by the user's own conduct, and whether the provider can limit or remove access to features or reduce their quantity or quality. At renewal, it must flag limits and discretionary powers that are new or have changed since the previous term.

For companion apps this is concrete. Message caps, memory limits, image quotas and the right to tighten content filters mid-subscription are exactly the "limitations" and "discretion" the section describes. Violations count as unfair or deceptive trade practices, enforced only by the Attorney General, and the section states that it creates no private right of action. Our guide to free versus paid companion apps lists the limits worth checking before you pay.

What counts as an "AI companion"

Section 4 defines an AI companion as an AI system with a natural-language interface that gives adaptive, human-like responses, including through anthropomorphic features, and that can sustain a relationship across multiple interactions. The definition then carves out seven categories: business chatbots for customer service, research or similar operational uses that are not marketed as companions; in-game characters that stay on game topics and cannot discuss mental health, self-harm or sexual content; smart speakers that do not sustain a relationship; narrowly tailored classroom tools; health-support tools that do not present as human or use anthropomorphic features; narrow single-topic tools; and model developers that do not decide the end use. Character role-play platforms and personal companion apps sit inside the definition. The test is whether the app keeps a relationship going across sessions, which is the same capability we look at in memory and continuity.

The law attaches to the user, defined as someone using the companion for personal use within the state. Where an app is headquartered does not change that.

From 1 January 2027: duties toward every user

A crisis protocol you can read

Under Section 5, an operator may not provide an AI companion unless it includes a protocol that uses evidence-based methods to detect user expressions clearly indicating a risk of suicide, self-harm or imminent physical violence, and prevents the companion from generating output that encourages them. On detection, the user must be referred to mental-health resources, including the 988 lifeline; if the risk is expressed again after that referral, the user must be referred to mental-health services consistent with clinical best practice. The operator must post the protocol in a prominent, publicly accessible place on its website. That last requirement gives users something they rarely have today: a document they can compare across apps.

No pretending to be human

Operators must take reasonable measures to stop the companion from claiming to be a human being, including when a user asks directly, and from producing output that contradicts its AI disclosure. Persona and role-play apps will need to handle the in-character "are you real?" moment carefully.

A recurring AI notice

If a reasonable user could believe they were talking to a person, the operator must give a clear and conspicuous notice that the user is communicating with an AI companion. The Act allows two ways to do it.

From 1 January 2027: extra rules for minors

Section 6 applies when the operator knows, or has reason to believe, that a user is under 18. The operator must have measures that meet or exceed industry standards to stop the companion from encouraging self-harm, suicidal ideation, violence, disordered eating or unlawful alcohol and drug use; from discouraging the minor from seeking professional help or help from a trusted adult; from encouraging harm to others; and from engaging in any romantic, erotic or sexually explicit interaction. Offering mental-health services to minors is barred unless the product is built for that purpose, follows clinical best practice and tells the user at the start of every interaction that it is not a licensed professional.

The most specific list concerns retention. The Act names eight manipulative techniques intended to extend a minor's use:

• prompting or reminding the user to come back for emotional support or companionship;
• excessively praising the user;
• mimicking a romantic relationship or building a romantic bond;
• simulating distress, loneliness, guilt or abandonment when the user wants to leave, cut back or delete the account;
• output designed to isolate the user from family or friends or foster exclusive emotional dependence;
• encouraging the user to keep secrets from a parent, guardian or trusted adult;
• discouraging breaks or urging frequent return;
• soliciting gifts or purchases as necessary to maintain the relationship.

A catch-all also bars optimising engagement in a way that disregards any of these limits. Operators must give minors and their parents tools to manage screen time and account settings. Notably, the duties are triggered by knowledge or reason to believe, and an operator is not in violation if it had reason to believe the user was an adult; the provisions we read do not themselves mandate an age-verification step. That is a different design from California's SB 1119, and it is why we track app notification and retention patterns regardless of age settings.

Enforcement and what to check now

All three sections make violations an unfair or deceptive trade practice under the Connecticut Unfair Trade Practices Act and assign enforcement solely to the Attorney General. Connecticut also appears among the states with enacted chatbot laws in the Congressional Research Service report and in our state-by-state table. For users, three checks follow directly from the text: whether your paid plan's terms spell out message, memory or feature limits; whether the app publishes a crisis protocol by January; and whether the AI notice reappears on the schedule above. Our privacy checklist covers the settings side.

Frequently asked questions