Connecticut Now Says Expressly That AI Is No Defense to Employment Discrimination. Employers With One Employee Can Be Covered. Written Notices Follow in October 2027.

1. What changed for Connecticut employers on October 1, 2026?

On October 1, 2026, Connecticut began saying expressly that an employer’s use of an AI hiring or HR tool is not a defense to a discrimination complaint, and it added an AI question to WARN layoff notices. Both changes come from Public Act 26-15, Substitute Senate Bill 5, “An Act Concerning Online Safety.” The Senate passed it on April 21, 2026, the House on May 1, 2026, and Governor Lamont signed it on May 27, 2026, according to the General Assembly’s bill history and the “Approved May 27, 2026” line at the end of the Act itself. Some summaries give June 2; the primary record says May 27.

Public Act 26-15 is a 74 page omnibus that also regulates chatbots, AI companions, frontier developers and youth social media. Employers need four parts of it: Sections 7 to 12 (definitions, vendor duties and the 2027 notice regime), Section 13 (the amendment to the main employment discrimination statute), Section 14 (the parallel amendment for sexual orientation and civil union status) and Section 26 (the WARN disclosure). Law firms and the press have started calling the whole package the CART Act.

Here is the short version for an employer. Connecticut Public Act 26-15 makes the use of automated employment related decision technology no defense to a discrimination complaint as of October 1, 2026, requires WARN filers to say whether layoffs relate to AI, and from October 1, 2027 requires plain language and written notices to applicants and employees, enforced by the Attorney General.

Those dates create two clocks. The first, already running, is about liability: if a claimant proves discrimination in a decision a tool shaped, the tool will not carry the blame for you. The second, eleven months away, is about transparency: telling people the tool exists and what it uses. Most commentary has focused on the 2027 notices. The 2026 change matters more today, because complaints about decisions made this month are already within the filing window.

2. Which employers does Connecticut’s AI employment law reach?

The October 2026 discrimination change reaches every employer covered by the Connecticut Fair Employment Practices Act, and under Conn. Gen. Stat. 46a-51(10) that means any person or employer “with one or more persons” in its employ, plus the state and its political subdivisions. There is no threshold above one employee, no revenue threshold and no general exemption for small firms. The definition of “employee” in 46a-51(9) does exclude an individual employed by his or her parents, spouse or child, so a business staffed only by family members in those relationships may sit outside it. A three person agency with ordinary staff that screens résumés through an AI matching feature is inside the rule.

The scale matters. The SBA Office of Advocacy’s 2025 Connecticut profile counts 381,129 small businesses, of which 68,418 are employer firms, and 726,097 small business employees, 48.1 percent of Connecticut employees. The profile uses 2022 Census data and defines small businesses as firms with fewer than 500 employees, so these are statistical counts, not a census of legally covered employers. The one employee threshold means even very small employers in that population may be covered when they use a covered tool, subject to the statute’s definitions and other applicable limits.

The 2027 notice duties use a different trigger. They apply to a “deployer,” defined in Section 7 as a person “doing business in the state” who deploys a covered tool in the state, and the notices protect an “employee or applicant for employment in the state.” An out of state company hiring remote workers who live in Connecticut should assume those notices will apply to it.

The WARN disclosure in Section 26 is the narrowest piece. It only applies to an employer that serves written notice on the Labor Department under the federal WARN Act. The Connecticut Department of Labor describes that law in general terms as reaching employers with 100 or more full time workers facing a plant closing or mass layoff, but federal coverage turns on its own counting rules and exceptions.

3. What counts as automated employment related decision technology?

Automated employment related decision technology is any technology that processes personal data and uses computation to generate an output, such as a prediction, recommendation, classification, ranking or score, that is a substantial factor used to make or materially influence an employment related decision. That is the Section 7 definition, and it does not use the words “artificial intelligence” at all. A rules based scoring script can qualify if it meaningfully changes outcomes.

“Substantial factor” has its own definition: a factor, including a constraint, ranking, score, recommendation or classification, that “meaningfully alters the outcome” of an employment related decision. A knockout filter that removes applicants without a degree is a constraint. A match score that decides which 20 of 300 résumés a recruiter opens is a ranking. A ranking or recommendation is not covered merely because it exists; it must meet the substantial factor test. Both of these examples can meaningfully alter outcomes, and a person making the final call does not by itself take a tool outside the definition.

“Employment related decision” covers decisions based on an individual’s personal data to hire, promote, discipline, discharge, renew, select for training or apprenticeship, or affect tenure or terms, privileges or conditions of employment. It excludes decisions that cause only a nonmaterial change in job tasks, responsibilities, hours or assignments, and decisions about workplace health and safety, scheduling and planning, or productivity monitoring. Personal data has the meaning in the Connecticut Data Privacy Act at Conn. Gen. Stat. 42-515.

Three exclusions keep ordinary software out: word processing, spreadsheets, databases, firewalls, spam filtering and similar tools “insofar as” they do not make or materially influence a decision; any system used in a manner “incidental” to a decision; and information that is purely descriptive, diagnostic or statistical and not relied on to make or influence the decision.

The table is our reading of the text, not guidance from the Attorney General or the Commission, neither of which had issued interpretive material on these definitions as of October 22, 2026, so far as we found. The last row is where small employers most often get caught: the exclusion attaches to the decision type, not to the product category.

4. What does “not a defense” mean in a discrimination complaint?

It means an employer cannot answer a Connecticut discrimination complaint by saying the AI tool made or shaped the decision. Section 13 adds this sentence to Conn. Gen. Stat. 46a-60(b)(1): the use of an automated employment related decision technology “shall not be a defense against a complaint alleging a discriminatory practice in violation of this subdivision.” Section 14 adds the same rule to 46a-81c for sexual orientation and civil union status.

The sentence does not create a new protected class, a new cause of action or a new burden of proof. It expressly forecloses reliance on the technology itself as a defense. It does not establish that outsourcing a decision to an algorithm previously immunized an employer; existing discrimination law already reached employer practices involving automated tools, as the litigation discussed below shows. The practical effect is that an employer cannot avoid an otherwise established discrimination claim merely by pointing to the tool or the vendor. The claimant must still prove the underlying statutory violation, and other legally available defenses remain.

The same subsection then tells the Commission on Human Rights and Opportunities and the courts what they may weigh instead. They may consider evidence of anti bias testing or similar proactive efforts to avoid the discriminatory practice, including “the quality, efficacy, recency and scope” of that testing or those efforts, “the results” and “the response thereto.” Four qualities, the results, and what you did about them. The provision permits that evidence to be considered; it does not itself require any particular testing program, frequency or statistical method.

There is a drafting detail staffing firms should notice. The new sentence applies to subdivision (1) of 46a-60(b), which governs an “employer.” The neighboring subdivision (2), which governs an “employment agency” that fails to classify or refer people properly, was not given the same sentence. A staffing firm acting as the employer of its temporary workers is squarely in subdivision (1). Whether a firm acting purely as a referral agency can still raise its tool as a defense is an open question we would not build a compliance strategy on, and the 2027 notice duties apply to any deployer regardless.

The national litigation shows that automated screening already carried discrimination exposure before Connecticut acted. In Mobley v. Workday, Case No. 3:23-cv-00770, the Northern District of California preliminarily certified a nationwide age discrimination collective on May 16, 2025 over an AI applicant recommendation system, finding the plaintiffs had plausibly alleged a “unified policy” of using the tool to score, sort, rank or screen applicants. According to Proskauer’s summary, Workday represented in filings that 1.1 billion applications were rejected using its software tools during the relevant period. That figure counts applications, not unique applicants, and is not a judicial finding that those applications were rejected discriminatorily or because of AI. The certification was preliminary, not a merits finding, and the case targets the vendor, but the hiring decisions involved were made by employers.

Remember the timing. Under Conn. Gen. Stat. 46a-82(f), a complaint for discrimination occurring on or after October 1, 2021 must be filed with the Commission within 300 days. A screening decision made in October 2026 can be the subject of a complaint until roughly August 2027, before the notice regime even starts.

5. What must employers disclose about AI in layoffs?

Since October 1, 2026, each employer that serves written notice on the Connecticut Labor Department under the federal WARN Act must also disclose “whether the layoffs that are the subject of such written notice are related to the employer’s use of artificial intelligence or another technological change.” That is the full text of Section 26, apart from a cross reference to the Act’s general AI definition. The disclosure goes to the Department “in a form and manner prescribed by the Labor Commissioner.”

The trigger is the federal one. The Department of Labor’s WARN page describes it as an employer with 100 or more full time workers that, within a 30 day period, closes a site where 50 or more workers lose employment, or lays off 50 to 499 workers who are at least 33 percent of a site’s workforce, or lays off 500 or more. Most small businesses will never file. Whether a particular closure or layoff triggers federal notice requires a separate analysis of employment losses, aggregation and exceptions; the 100 worker shorthand is not the complete test.

Two points are easy to miss. First, the question covers AI “or another technological change,” so an automation project without any machine learning still has to be considered. Second, the statute asks whether the layoffs are “related to” that use, a broader test than whether a model selected who was laid off. Replacing a call center team with a chatbot is related; so, arguably, is a reorganization justified by expected AI productivity gains.

As of October 22, 2026, the Department’s WARN page we reviewed did not show a separate form for the AI disclosure. Confirm the Department’s current filing instructions before any filing. If no separate mechanism is specified, include an express AI or technology statement with the WARN submission and retain the supporting analysis, recognizing that later instructions may require a different format. California took a parallel step on September 30, 2026 with SB 951’s technology displacement notices, so multistate employers should build one internal process for both.

6. What notices are required from October 1, 2027?

Starting October 1, 2027, Connecticut requires two notices from employers that deploy covered tools: a plain language interaction disclosure and a written pre decision notice. Sections 9 and 10 took effect as law on October 1, 2026, but each applies to a deployer who, “on or after October 1, 2027, deploys” the technology.

The interaction disclosure (Section 9). If a covered tool is intended to interact with an employee or applicant, such as an AI interviewer, a screening chatbot or an assessment game, the deployer must disclose in plain language that the person is interacting with it. No disclosure is needed where “a reasonable person would deem it obvious.”

The written pre decision notice (Section 10). If a deployer uses a covered tool to make, or as a substantial factor in making, an employment related decision, it must give the employee or applicant a written notice, before the decision is made, disclosing six items:

1. that the deployer has deployed such a technology;
2. the technology’s purpose and the nature of the decision;
3. the trade name of the technology;
4. the categories of personal data it will analyze or process, and how that data will be assessed in reaching a decision;
5. the sources of that personal data; and
6. contact information for the deployer.

Item 4 is the hard one. “How the personal data will be assessed” requires an employer to explain, at least at a general level, what the tool does with a résumé or interview. Few small employers can write that sentence today without information from the vendor, which is why Section 8 exists.

Section 11 lets a person withhold information that is a trade secret or otherwise legally protected, but it must then send a notice saying it is withholding information and the basis for doing so. That provision protects vendors’ model details; it does not excuse an employer from giving the notice itself.

One interpretive question deserves attention now. “Deploy” means “to put” a tool “into use.” A tool first switched on before October 1, 2027 and still running afterward is arguably not newly deployed, and some summaries describe the duties as applying to technology deployed on or after that date. The statute’s words are “deploys on or after,” and we do not present continued use as a settled interpretation. Our recommendation is practical rather than legal: plan notices for every covered tool still in use on October 1, 2027 instead of relying on an earlier activation date. No guidance resolves this yet.

7. What do AI vendors owe employers under the new law?

Developers of covered tools deployed in Connecticut on or after October 1, 2027 must give deployers “all information” the deployer needs to perform its Section 9 and 10 notice duties. A developer is anyone doing business in the state who develops, or intentionally and substantially modifies, a covered tool.

The duty has a scope limit in Section 8(b). A developer owes nothing unless the tool was “advertised, marketed, configured, contracted for, sold or licensed” to be used to materially influence an employment related decision. A general purpose chatbot used for screening without being advertised, configured, contracted for, sold or licensed for that purpose may fall outside the developer information duty trigger. The employer’s own notice obligations require separate analysis, and in that case it may have to assemble the notice content without the developer’s help.

Section 8(c) allows a developer to contract with a deployer to assume the deployer’s notice duties. The contract must be binding and must “clearly set forth” which duties the developer has assumed. For a small employer that uses one applicant tracking system for every hire, this is a practical compliance route: ask the vendor to deliver the notices inside its candidate workflow, in a contract that names Sections 9 and 10. A contract may allocate specified notice duties to the developer, but the employer should retain evidence that the agreed notices are actually delivered and obtain advice on any remaining obligations.

Do not confuse that with liability transfer. Section 8(c) addresses notice duties. It does nothing to Section 13. If a discrimination complaint is brought over a decision the tool shaped, the employer is still the party answering it and its use of the vendor’s tool is still not a defense, though the claimant must prove the claim and other defenses remain. Indemnity language can share the cost of that risk; it cannot remove the employer from the case.

8. Who enforces Connecticut’s AI employment rules and what are the penalties?

Two different enforcement systems apply, and they should not be confused. The October 2026 discrimination rule is enforced the way all Connecticut employment discrimination claims are: through complaints to the Commission on Human Rights and Opportunities and, after that process, in court. The 2027 notice and vendor rules are enforced only by the Attorney General.

Section 12 makes any violation of Sections 8 to 11 an unfair or deceptive trade practice under the Connecticut Unfair Trade Practices Act, enforced “solely by the Attorney General.” It states that the CUTPA private action at Section 42-110g does not apply and that nothing in Sections 8 to 11 creates a private right of action. An applicant who did not receive a notice cannot sue for that omission alone.

Under Conn. Gen. Stat. 42-110o, a court may impose a civil penalty of up to $5,000 for each wilful violation, where the violator “knew or should have known” its conduct violated CUTPA, and up to $25,000 per violation of a restraining order or injunction. Because the notice duty runs to each applicant and employee, exposure can scale with headcount, though how violations will be counted under this statute is untested.

For violations occurring on or before December 31, 2027, the Attorney General “may” issue a notice of violation before suing if the violation can be cured, and the employer then has 60 days to cure. That window is discretionary and short. For violations after December 31, 2027, this provision no longer supplies the special discretionary 60 day cure process, although that does not prevent voluntary correction or settlement.

9. How does Connecticut compare with Illinois, Colorado, California and New York City?

Among these five regimes, only Connecticut’s text says in express terms that using the tool is not a defense; the others focus on notices, audits, effects based liability or human review. The table compares the core employer duties as enacted.

For Colorado, see our breakdown of what actually applies to employers on January 1, 2027. For California, our SB 947 analysis walks through the enrolled text; the Governor signed it on September 30, 2026, and it becomes operative July 1, 2027. Illinois is also worth reading alongside our piece on its 2026 disparate impact amendments.

The practical lesson is that the controls overlap. One inventory, one vendor information request and a shared testing and notice program can reduce duplication across all five regimes. They do not make the regimes interchangeable: they differ in scope, timing, audit independence, public disclosure and individual rights, and a general testing practice will not necessarily satisfy New York City’s specific bias audit requirements.

10. What should a small Connecticut employer do in the next 90 days?

A small Connecticut employer should spend the next 90 days building the evidence that Section 13 lets a fact finder weigh: an inventory, a test, and a documented response. The 2027 notices can be drafted from the same work. Adoption is real but uneven: in SHRM’s State of AI in HR 2026 research, drawn from more than 1,900 HR professionals, 39 percent of organizations had implemented AI in their HR functions, and SHRM reports recruiting as the area where that use is most concentrated, at 27 percent.

Days 1 to 30: find every tool

List every system that touches hiring, promotion, discipline, discharge, training selection or pay: the applicant tracking system, job board screening features, assessment and interview platforms, background check scoring, performance software, and any generative AI a manager uses to compare candidates. For each one, record whether it ranks, scores, filters or recommends people, and which decision it feeds. Check the vendor’s release notes, because AI features are often added to existing subscriptions. Mark the excluded uses (scheduling, safety, pure monitoring) and write down why.

Days 31 to 60: test and respond

Ask each vendor in writing for its most recent bias testing, the date, the population tested and the method. Where your own volume allows, compare selection rates by sex, race and age band at each stage the tool controls, and document the result and what you changed. If volume is too small for statistics, document human review of the tool’s rejections and a reasoned check of its inputs for proxies such as zip code or graduation year. These are our recommended controls, not statutory requirements. Date the testing and document your response. Establish a risk based review schedule and reassess after material changes; annual review is one possible baseline, not a statutory frequency.

Days 61 to 90: contracts, WARN and notices

Send each in scope vendor a Section 8 information request covering trade name, data categories, data sources and how data is assessed. Ask whether it will assume the Section 9 and 10 notices under Section 8(c) and get that in a contract. Add one line to your layoff playbook so any WARN filing includes the Section 26 AI answer. Draft the six item notice now, so October 2027 is an update rather than a project. An ISO/IEC 42001 aligned management system or the NIST AI RMF gives this work a structure, though alignment is supporting evidence of a sound program, not a substitute for testing your own tools.

This article is provided for general informational purposes and reflects the state of the law as of October 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.