When does Connecticut's AI law take effect?

In stages. October 1, 2026 for subscription disclosures, frontier whistleblower rules, provenance data, the WARN question and the "not a defense" rule. January 1, 2027 for AI companion chatbots. October 1, 2027 for notices to applicants and employees about automated employment decision tools.

What does SB 5 require for AI companion chatbots?

From January 1, 2027, a suicide and self-harm protocol posted on the operator's website, no claims to be human, and a notice that the user is talking to an AI. The notice is either always visible or repeated, at least hourly for minors and every three hours for adults.

Who must add provenance data to AI-generated content?

Only covered providers, meaning companies whose generative image, audio or video system has more than one million monthly users and is open to consumers for personal use. Since October 1, 2026 they must include provenance data where commercially and technically reasonable. Business-to-business use is excluded.

Who enforces Connecticut's AI law?

The Attorney General, as unfair trade practices. The employment notice, companion, provenance and subscription sections say enforcement is by the Attorney General alone, and several state there is no private right of action. Frontier whistleblower violations carry civil penalties up to $1,000 each.

Connecticut AI Law (SB 5): Requirements and Checklist

Q: What is an automated employment-related decision technology under SB 5?

Technology that processes personal data and uses computation to produce an output that is a substantial factor in an employment decision, meaning it meaningfully alters the outcome. Word processing, spreadsheets and spam filters are excluded, and so are decisions about scheduling, workplace safety and productivity monitoring.

The US Capitol at the end of a lawn with autumn trees, used here as a generic image of a legislature Image: Pexels, used under the Pexels License.

If you hire in Connecticut, the part of SB 5 that already applies to you is one sentence: since October 1, 2026, using an automated hiring tool is not a defense to a discrimination complaint. The notices most coverage talks about start a year later, on October 1, 2027. If you run a companion-style chatbot, your date is January 1, 2027. Everything else in the act lands on AI vendors, big image generators and frontier labs.

The bill passed the Senate 32-4 on April 21 and the House 131-17 on May 1, as CT Mirror reported. It became Public Act 26-15, titled "An Act Concerning Online Safety", and its signature line reads "Approved May 27, 2026". The other dates in law firm summaries come from the Governor's office: its bill notification of May 29 lists SB 5 among the bills signed "on May 27", and Governor Ned Lamont announced the signing at a June 2 event. So May 27 is the legal date; June 2 is the ceremony. Law firms also call it the CART Act (Connecticut Artificial Intelligence Responsibility and Transparency Act), a name that doesn't appear in the 74-page act.

Video: "Bill signing for legislation establishing youth online safety protections and AI regulations", Governor Ned Lamont's YouTube channel.

What applies when

Date	What starts	Who it's for	Sections
October 1, 2026	Written key-terms notice before an AI subscription starts or renews	Subscription AI providers selling to Connecticut residents	Sec. 1
October 1, 2026	Whistleblower protections; anonymous internal reporting by January 1, 2027 for the largest labs	Frontier developers (over 10^26 operations); large ones over $500 million revenue	Sec. 2
October 1, 2026	Using an automated employment tool is not a defense to a discrimination complaint	Employers	Secs. 13, 14
October 1, 2026	Provenance data in generated audio, images and video	Generators with over 1 million monthly users, open to consumers	Sec. 15
October 1, 2026	Tell the Labor Department whether layoffs in a federal WARN notice relate to AI	Employers filing WARN notices	Sec. 26
January 1, 2027	Safety protocol, human-impersonation ban, AI notice, extra rules for minors	AI companion operators	Secs. 4-6
October 1, 2027	Disclosure and six-item pre-decision notice	Deployers of automated employment decision tools, and their developers	Secs. 8-10

Video: WTNH News8, "Connecticut Senate passes significantly amended AI regulation bill".

Hiring tools: one rule now, notices in 2027

The act defines an automated employment-related decision technology as anything that processes personal data and uses computation to produce an output that is "a substantial factor" in an employment decision, meaning it meaningfully alters the outcome. Word processors, spreadsheets and spam filters are out. So are decisions about scheduling, workplace safety and productivity monitoring, which the act carves out of "employment-related decision" by name. If an earlier guide told you your shift-scheduling tool is covered, it isn't.

What applies today is Sec. 13: "The use of an automated employment-related decision technology... shall not be a defense against a complaint alleging a discriminatory practice." The commission or a court may weigh your anti-bias testing, including its quality, recency, scope and results. That's not an audit mandate. It's a reason to have the file anyway.

From October 1, 2027, a deployer must disclose in plain language when an applicant or employee is interacting with such a tool, unless that's obvious, and must give a written notice before a decision with six items: that the tool is used, its purpose and the decision, its trade name, the categories of personal data and how they're assessed, the data sources, and contact details. Developers must hand deployers what they need for those notices if the tool was marketed or sold to influence employment decisions, and a developer can take the duties over by contract.

The act has no right to correct data, no appeal and no annual bias audit. Our Connecticut SB 5 employer checklist walks through the hiring sections step by step.

AI companion chatbots: January 1, 2027

Sections 4 to 6 cover "artificial intelligence companions": systems with a natural language interface that give human-like responses and can sustain a relationship across interactions. A customer-service or internal productivity bot that isn't marketed as a companion is excluded, as are narrow task tools, in-game bots and classroom tools.

Page 12 of Connecticut Public Act 26-15 showing Section 5(b), the AI companion notice timing, and the start of Section 6 on minors Screenshot: Public Act 26-15, page 12 (Secs. 5 and 6), Connecticut General Assembly.

If you operate one, you need:

A protocol that detects expressions of suicide, self-harm or imminent violence, refers the user to resources including the 988 lifeline, and is posted on your website.
Reasonable measures so the companion never claims to be human, even when asked.
A clear notice that the user is talking to an AI, if a reasonable user could think otherwise. Either keep it visible for the whole session, or show it at the first interaction in any 24-hour period and then at least hourly for users under 18, or every three hours for adults.
For users you know or have reason to believe are minors: measures that stop the companion encouraging self-harm, violence, disordered eating or drug and alcohol use, any romantic or sexual interaction, and eight listed manipulation tactics, plus screen-time and account tools for minors and parents.

Enforcement is by the Attorney General alone, as an unfair trade practice, and Sec. 5 states there's no private right of action. The earlier version of this guide said parents could sue directly for punitive damages. The enacted text doesn't say that.

Provenance data for big generators: October 1, 2026

Page 28 of Connecticut Public Act 26-15 showing Section 15(b): provenance data duties and the reference to the Coalition for Content Provenance and Authenticity standard Screenshot: Public Act 26-15, page 28 (Sec. 15), Connecticut General Assembly.

Sec. 15 applies to "covered providers": whoever produces a generative system for images, audio or video that has more than one million users a month and is publicly accessible to consumers for personal use. To the extent commercially and technically reasonable, they must include provenance data that lets a consumer check whether content was created or materially altered by the system, and make it hard to strip, with the C2PA standard named as one method. Business-to-business use, video game experiences and pure upscaling or noise-reduction tools are excluded.

There is no second-stage watermarking rule for "all developers" in October 2027. That was in drafts we relied on earlier and isn't in the enacted act. If you only use these tools for marketing, the duty sits with your vendor.

Subscriptions, frontier labs and WARN notices

Subscription AI (Sec. 1). A provider can't start or renew a paid AI subscription with a Connecticut consumer unless it has given a written notice of key terms (usage limits, and whether it can cut or reduce functionality) and the consumer has accepted in writing.

Frontier developers (Sec. 2). A frontier developer trained a foundation model with more than 10^26 operations. It can't stop staff from reporting catastrophic-risk concerns and must post or send a notice of their rights. A large frontier developer (over $500 million in annual revenue) needs an anonymous internal reporting process by January 1, 2027. Penalties run up to $1,000 per violation, collected by the Attorney General.

Page 44 of Connecticut Public Act 26-15 showing Section 26: employers serving WARN notices must tell the Labor Department whether the layoffs relate to artificial intelligence Screenshot: Public Act 26-15, page 44 (Sec. 26), Connecticut General Assembly.

WARN notices (Sec. 26). An employer that serves a notice on the Connecticut Labor Department under the federal WARN Act (29 USC 2102(a)) must say whether the layoffs relate to its use of AI or another technological change, in a form the Labor Commissioner sets. It rides on the federal threshold (employers with 100 or more employees). We found no separate penalty in Sec. 26.

Checklist by date

Now (in force since October 1, 2026):

List every tool that scores, ranks or screens Connecticut applicants or staff, and keep its anti-bias testing results on file.
Add an "AI or other technological change?" field to your WARN notice template.
If you sell AI subscriptions to Connecticut consumers, put the key-terms notice and written acceptance into checkout and renewal.
If you buy image, audio or video generation, ask the vendor whether it's a covered provider and how it embeds provenance data.

Before January 1, 2027:

Decide whether your chatbot is a companion under Sec. 4 or falls in an exclusion, and write down why.
If it's a companion: publish the safety protocol, block "I'm human" answers, and set the notice to always-on or hourly for minors and every three hours for adults.

Before October 1, 2027:

Draft the plain-language disclosure and the six-item pre-decision notice.
Get your hiring-tool vendor's information for those notices in writing, or a contract clause where the vendor takes the notices over.

Our take

We'd spend the first hour on the hiring-tool inventory and the bias file, because that's the one piece already live for ordinary employers and it's what the act tells the commission or a court it may weigh. The companion rules matter only if you sell something people talk to for company. For everyone else, SB 5 is mostly a set of questions to put to vendors.

How we checked this: on October 13, 2026 we read Public Act 26-15 from the Connecticut General Assembly PDF (Secs. 1, 2, 4 to 15 and 26) and took the three screenshots above, checked the vote dates on the legislature's bill status page, and compared the votes and signing date with CT Mirror and the Ropes & Gray and Ogletree summaries. On October 15, 2026 we matched the signing date against the Governor's bill notification and June 2 press release. Last reviewed: October 13, 2026.

Correction (October 13, 2026): Earlier versions of this guide said the employment notices, data correction rights and an all-developer watermarking rule apply from 2026 or 2027, that scheduling tools are covered, that parents can sue over companion chatbots with no cure period, that the AG can impose $5,000 per violation, and that the bill passed on May 11. None of that matches the enacted act. The dates and duties above follow Public Act 26-15; May 11 is the day it got its Public Act number.