Executive Summary

Key Numbers

Source: Connecticut Public Act 26-15 (Substitute Senate Bill No. 5), compiled from the full 74 pages of the original PDF.

The act Connecticut passed this year is titled An Act Concerning Online Safety, and it is a large one: 39 sections across 74 pages. It reaches online safety, health care, employment and the labeling of generative AI content all at once. Section 26, the piece that takes effect on October 1, reads in full as follows.

"Each employer that serves written notice on the Labor Department pursuant to 29 USC 2102(a), as amended from time to time, shall disclose to the department, in a form and manner prescribed by the Labor Commissioner, 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."

Connecticut Public Act 26-15, Section 26 (effective October 1, 2026)

One sentence follows the one quoted above, and it says that artificial intelligence here carries the meaning given in Section 17 of the same act. After that the text ends.

It helps to settle first who is covered, because this is not a clause that creates a new filing. It adds a line to a filing that already existed. The federal WARN Act requires 60 days' notice from a company with 100 or more employees, part-timers excluded, when 50 or more workers at one site lose their jobs within 30 days and that group is at least 33 percent of the site's workforce, or when the number reaches 500, or when a site closes and 50 or more workers lose their jobs. Section 26 asks for one more answer on the paper that already goes in.

1.1The Much Wider Word the Statute Chose

Several summaries of this clause render it as a duty to disclose whether AI contributed to the layoffs. The words the statute actually uses are "related to," which is far wider than contribution or cause. Even with no AI anywhere near the decision to cut, a year in which the budget moved toward AI equipment can be related. The threshold was set low, and a low threshold means more companies land inside it.

A low threshold sounds like it should make the answer easier. It works the other way around. A question about cause has a place to check: was an AI output used in making this decision? A question about relation has no such place. How far back to go, and what counts as related once you get there, is left to the company writing the answer.

1.2Three Blanks Still Open

Taken one word at a time, the text leaves three spaces that nobody has filled yet.

• • The form. All the text says is "in a form and manner prescribed by the Labor Commissioner." Whether the answer is a checkbox, a written explanation, or a filing that carries supporting documents sits outside the text. Filing a WARN notice in Connecticut today does not involve a form at all. State labor department guidance says written notification should be on company letterhead, signed and dated by the authorized employer representative, and either emailed or mailed to the Rapid Response Unit, and it lists four things that notification should carry: the name and address of the impacted employment site, the date or dates of the proposed closing or mass layoff, the number of affected workers along with the address of their collective bargaining representative and chief elected officer where applicable, and the name, address and telephone number of the employer representative to contact. As of September 28, 2026, AI appears nowhere in that guidance, and no separate standard form is publicly available.
• • What "another technological change" means. The phrase turns up once in the 74-page act, in this paragraph, and no definition section covers it. Warehouse automation is a technological change. So is a newly installed enterprise system, and so is the last step of a 20-year office computerization project.
• • The basis for the judgment. The text does not say what a company should look at in deciding whether the relation holds. There is no duty to create a record and no duty to keep material that would support the answer.

The definition of artificial intelligence is the one space that is not empty. Section 26 borrows it from Section 17, where it covers any machine-based system that, for an explicit or implicit objective, infers from the inputs it receives how to generate outputs such as content, decisions, predictions or recommendations that can influence physical or virtual environments. That broad wording is shared across several state AI laws. Where the definition sits is worth a look, though. Section 17 is the clause that has the state establish a Connecticut AI Academy, an online course program, and the definition binds its own scope with the words "as used in this section." Borrowing a definition from another section is ordinary legislative practice. Still, the word that will decide layoff filings arrived from the glossary of a clause about online coursework, which says something about how quickly this duty was added.

Pebblous covered this wave in July, alongside California's moves, in the new U.S. labor rules turning AI layoffs into a disclosure field. The question left open there was who defines the label "AI-driven layoff," and by what standard. This article is about what comes next. Even once the definition is settled, is there anything inside the company to open when the answer has to be written down?

A good many write-ups of this law say that failing the AI disclosure becomes an unfair or deceptive trade practice enforced by the state Attorney General. In the full text, that sentence is attached somewhere else.

The provision in question is Section 12. A violation of Sections 8 to 11 constitutes an unfair or deceptive trade practice under subsection (a) of section 42-110b of the general statutes and is enforced solely by the Attorney General; for violations occurring on or before December 31, 2027, the Attorney General may first issue a notice of violation and allow 60 days to cure where the violation is judged curable; and nothing in those sections creates a private right of action. The scope written into the text is "sections 8 to 11, inclusive."

Sections 8 through 11 deal with automated employment-related decision technology. They set duties for the parties who develop and deploy tools that make employment decisions such as hiring, promotion and evaluation, or that substantially influence them. A developer has to hand the deployer the information the deployer needs to meet its own duties, and a deployer has to tell employees and applicants who interact with the tool, in plain language, that they are doing so. Those duties take practical effect on October 1, 2027. Different provision, different parties, different date from Section 26.

Section 26 carries no language about what happens if the duty is not met. Nothing else in the 74 pages mentions Section 26. How a breach of this duty would be handled under the general doctrine of the Connecticut Unfair Trade Practices Act is an open question, and not one to settle here. What the text itself shows is clear enough. The duty to write it down is there. The consequence of leaving it blank, or of getting it wrong, is not.

2.1Where the Law Calls Evidence by Name

The interesting part comes here. The same law has already written down once what evidence is. Sections 13 and 14 amend the anti-discrimination statutes so that using an automated employment-related decision technology cannot serve as a defense against a discrimination complaint. Then they add this.

"The commission or court may consider evidence of anti-bias testing or similar proactive efforts to avoid such discriminatory practice, including, but not limited to, the quality, efficacy, recency and scope of such testing or efforts, the results of such testing or efforts and the response thereto."

Connecticut Public Act 26-15, Sections 13 and 14 (effective October 1, 2026)

Six properties of a record sit in that one sentence. Was the testing done properly, did it actually work, when was it done, how far did it look, what came out of it, and what did the company do about what came out. The law does not order a company to keep such records. It does name, one by one, what can help a company once a dispute starts.

Sections 13 and 14 take effect the same day as Section 26. Same law, same effective date, same area of employment, and a different posture toward documents. One side lists six properties of a record. The other takes down a yes or a no.

2.2Section 29 Orders a Study to Count These Layoffs

Move a little further down the numbering and Section 29 appears. It took effect on passage and is running already, and it directs the Institute for Municipal and Regional Policy at the University of Connecticut to study the impact of artificial intelligence on the state's workforce and to develop a comprehensive strategy for it. Two of the items the law names for that study overlap with the question this article is asking. One is an assessment of the methods available to track layoffs and job displacements in the state that are associated with artificial intelligence. The other is an assessment of the data elements the Labor Department and other state agencies already collect that could be used to understand and track that impact.

The strategy side is more direct still. It calls for recommendations on additional data elements the Labor Department and other agencies should collect, and for a framework for recurring analyses and public reporting. The report is due January 1, 2027. That is three months after Section 26 puts the new line on the notice.

Set the two clauses side by side and the shape gets strange. One hands the Labor Department a new data element. The other has a university find out which data elements the Labor Department holds. And Section 29 never mentions Section 26. The gap noted earlier, that no clause in the 74 pages refers back to Section 26, costs the most right here, because a study designed with knowledge of the new line would have listed its items differently.

From here on this is the reading of this article rather than what the statute establishes.

How hard it is to put the relation into a single line already shows in documents that are public. Microsoft said on July 6, 2026, that it was cutting roughly 4,800 jobs, 2.1 percent of its workforce. The note Chief People Officer Amy Coleman sent to employees is posted on the company blog, and two sentences in it stand next to each other.

"I also want to be direct that the roles eliminated today are not being replaced by AI."

"At the same time, what is true is that AI is changing how work gets done."

Amy Coleman, Chief People Officer, Microsoft, company blog, July 6, 2026

The two sentences do not contradict each other. A model taking over one position and a company deciding afresh where to put its money and its people are separate events. In the same note the company said it had redeployed more than 4,000 employees into new roles over the past year, including another 500 that month, and that more than 30 percent of eligible employees chose to take part in its recent voluntary retirement program. Microsoft is not a Connecticut case and this was not a filing made under this law. It is simply a recent instance of what it looks like when a company explains in its own words how a layoff and AI relate. Once that explanation has to fit on one line, people inside one company disagree about whether the answer is yes or no.

Cohen and Wolf, a Connecticut firm that walked local employers through this law, advised them to review their internal layoff policy, to draft guidance on "tracking the rationale for the layoffs," and to consider whether information supplied by AI fed into the layoff decisions. That is advice that treats this as a problem of building records rather than one of waiting for a form.

3.1Three Kinds of Document Behind One Answer

To answer that line, a company ends up having to lay three kinds of document side by side.

• • The adoption decision. The material that records why the AI was brought in. Investment approvals, executive memos, projected-benefit spreadsheets. Whether headcount reduction appeared as a number in those projected benefits is what matters in this pile.
• • Vendor output. Where what the tool actually did is left behind. The permitted-use clauses of the contract, logs stamped with deployment scope and timing, test results on performance and bias. Most of this is made outside the company and arrives from there.
• • The org design. The documents that record why this particular position went away. Role redesign proposals, redeployment reviews, headcount plans before and after the reorganization.

Leave any one of the three empty and the yes or no written on the notice stops being a finding and becomes an assertion. The difference shows up later. A finding has an original to walk back to and an assertion does not. Whether the question comes from a regulator, from a laid-off employee or from internal audit, what is needed the moment it arrives is not the conclusion but the place the conclusion came from.

3.2These Documents Were Written Long Before the Layoff

A timing problem sits on top of that. None of the three piles gets written after the layoff is decided. The adoption documents were typically drafted one to three years earlier, and the vendor contracts and deployment logs piled up around then. Nobody writing them then expected they would one day support a filing with the state labor department. What Section 26 starts asking about on October 1 is not the records a company will keep from now on. It is the condition of the records already sitting there.

This is also where the reason Sections 13 and 14 put recency on their list of evidence properties lands. A record is not graded only on whether it exists. When it was made is graded alongside. If nothing sits between an adoption document from three years ago and a reorganization plan from last month, whatever judgments happened in between live only in people's memories.

The reason a person was let go is becoming a field on an administrative document. Fields get tallied, and tallies eventually attract someone asking where the numbers came from. Connecticut is already getting ready for that tally. When the Section 29 report arrives on January 1, it will recommend what more the Labor Department ought to gather, and there is no reason the line that started collecting three months earlier would stay off the list. How much anyone can trust the resulting statistics on layoffs related to AI rests entirely on what each company opened before filling that line in. Writing about the integrity of the reason label in July came from the same concern.

In data quality work, the path a value traveled becomes the problem more often than the value itself. A wrong value can be corrected. A value whose procedure and whose hands are unknown cannot even be sorted into fix or leave alone. The new line in Section 26 becomes exactly that kind of value. What builds its path inside the company is the adoption decision, the vendor output and the org design, and those three live with finance, with procurement and legal, and with HR, each kept for a different purpose. Pulling one answer out of three piles that have never been in the same room is the work Section 26 actually assigns.

Korea has no such line yet. The question travels anyway. What place an AI tool adopted two years ago took in this year's reorganization, and whether a document explaining that still sits inside the company, can be checked before any law asks. If the one asking later is the board, a union or a reporter rather than a regulator, the paperwork needed is the same.

One last thought from the individual side. This clause will not tell me why I was pushed out. The notice goes to the state labor department, and the duty to inform individuals directly begins under a separate provision on October 1, 2027. What arrives on October 1 is a duty on companies to write, not a right for people to know. One thing does change, though. For the first time the reason gets set down in writing somewhere outside the company. Writing can be compared against other writing, and anything that can be compared can eventually be questioned.

Thank you for reading this far. Everything quoted here can be read in the text of Public Act 26-15 published by the Connecticut General Assembly. Pick one AI tool your own organization brought in over the past two years and go looking for where the document explaining why it was brought in and what it changed is kept today. We would like to hear what turns up.

R.1Official Documents & Legislation

• 1.Connecticut General Assembly. (2026). "Public Act No. 26-15 — An Act Concerning Online Safety (Substitute S.B. No. 5)."
• 2.Cornell Law School, Legal Information Institute. (2026). "29 U.S. Code § 2101 — Definitions." (Federal WARN Act).
• 3.Connecticut Department of Labor. (2025-08-18). "Rapid Response: WARN Act Notification."

R.2Industry & Company Cases

• 4.Ogletree Deakins. (2026). "New Connecticut Law Restricts Employer AI Use, Mandates Notice for AI-Caused RIFs."
• 5.Cohen and Wolf, P.C. (2026). "What Employers Should Know About Connecticut's New AI Laws."
• 6.Coleman, A. (2026-07-06). "The Latest in Our Company Transformation." Microsoft Official Blog.