Connecticut Governor Ned Lamont signed Senate Bill 5 — Public Act 26-15, the Artificial Intelligence Responsibility and Transparency Act, in the spring of 2026. This site published a full walk-through of the 39-section statute in June.

The Act’s obligations are phased from October 2026 through January 2028. The first tranche lands on 1 October 2026 — six weeks from the date of this article — and this piece is about that tranche specifically: what becomes enforceable on that date, and what has to be built before it.

A note on nomenclature before anything else, because it causes real confusion. SB 2 was the 2025 Connecticut AI bill. It passed the Senate and did not become law. The enacted statute is SB 5 / Public Act 26-15. Guidance and vendor materials still circulating that reference “Connecticut SB 2” are describing a bill that never took effect, and its provisions differ from what was enacted. Check the citation before relying on any analysis.

What takes effect on 1 October 2026

Four things. They hit different parts of an organisation, and in most companies no single person owns all four.

1. The developer–deployer allocation framework

The Act adopts the structure now familiar from Colorado and the EU AI Act: obligations are split between the party that develops an AI system and the party that deploys it, with documentation and information flowing from developer to deployer so the deployer can meet its own duties.

The practical consequence for most Connecticut employers is that they are deployers, and their ability to comply depends on receiving documentation from vendors. An HR technology vendor that will not provide the information a deployer needs is a compliance problem that cannot be solved internally.

The corollary matters too: many organisations are developers without thinking of themselves as one. A company that fine-tunes a model, builds a scoring tool on top of a foundation model, or substantially modifies a purchased system may sit on the developer side of the line for that system.

2. “AI is not a defence” — the anti-discrimination amendments

The Act amends Connecticut’s anti-discrimination statutes to provide that an employer may not avoid liability for unlawful discrimination by attributing the decision to an AI system — including where the technology was supplied by a third party.

This is the provision with the most immediate legal force, and it requires no implementation project at all. It simply removes an argument. As of 1 October 2026, “the vendor’s algorithm made that determination” is not a defence in a Connecticut discrimination claim.

What it does require is a change in how vendor relationships are papered. If the employer bears the liability regardless of who built the tool, then indemnification, audit rights, bias-testing evidence and the right to obtain the vendor’s validation documentation stop being nice-to-have contract terms and become the mechanism by which the employer manages an exposure it cannot otherwise transfer.

3. Automated employment decision technology notices

The Act creates a notice regime for automated employment-related decision technology (AEDT) used in hiring, promotion, discipline and other personnel decisions. Employers must tell workers and applicants when they are interacting with AEDT and provide written pre-decision notices.

The phasing here needs care, because it is the most commonly misread part of the statute. The framework takes effect on 1 October 2026. A further tranche of employment obligations applies to systems deployed on or after 1 October 2027. Organisations should not read the 2027 date as a blanket deferral of employment duties to 2027 — the 2026 tranche is live, and the anti-discrimination amendment above operates from 2026 regardless.

The word “interacting” is broader than it first reads. An applicant who never sees the tool still interacts with it in the relevant sense when it screens their application. Notice obligations attach to use, not to visibility.

4. Synthetic content provenance for large providers

From 1 October 2026, covered providers with more than one million monthly users must, to the extent commercially and technically reasonable, embed provenance data into AI-generated or materially altered audio, image or video content.

Two observations.

The “commercially and technically reasonable” qualifier is a genuine standard, not an escape hatch. It requires a documented assessment of what is feasible — and a provider that has never assessed feasibility cannot claim the limit applies.

The substantive requirement converges with EU AI Act Article 50. A provider building C2PA-style signed metadata and imperceptible watermarking to meet the EU’s Article 50 marking obligations is building most of what Connecticut asks for. Build once, map to both.

The one that surprises people: the AI layoff disclosure

Buried in the package is a provision that is drawing less attention than it deserves.

On and after 1 October 2026, an employer conducting a layoff substantially caused or contributed to by an artificial intelligence system must provide an AI-related layoff notice, alongside Connecticut’s existing separation and mass-layoff notice obligations.

This is, as far as we are aware, the first state requirement of its kind, and it is operationally awkward for reasons that have nothing to do with drafting.

It requires a causation determination that organisations do not currently make. When a company reduces headcount after deploying automation, who decides — and on what record — whether AI “substantially caused or contributed to” the reduction? That is a judgment with legal consequence, and in most organisations it currently sits nowhere. It is not an HR determination, not a finance determination, and not a legal determination, because nobody has ever had to make it.

It creates a documentation hazard in both directions. Internal materials that justify an automation investment on headcount-reduction grounds are exactly the evidence that would establish AI causation for a later layoff. Business cases, board decks and vendor ROI models routinely contain this language. Conversely, an organisation that determines AI was not a substantial contributor should be able to show the basis for that conclusion.

It interacts with existing WARN obligations, so the analysis has to happen on the WARN timeline, not afterwards.

The practical step is to establish, now, who makes the AI-causation call, on what evidence, and how it is recorded — before a reduction in force is being planned under time pressure and the question is asked for the first time in a room full of people who want a fast answer.

A six-week plan

Weeks 1–2: inventory and classify. Enumerate every AI or automated system touching employment decisions for Connecticut workers or applicants — including résumé screening and ranking in your ATS, video interview scoring, assessment tools, scheduling and work-allocation systems that affect compensation, and performance or discipline flagging. For each, determine whether you are developer, deployer, or both. Include tools embedded in HR platforms that were never procured as AI.

Week 3: vendor documentation. Write to every AEDT vendor requesting the developer-side documentation the Act contemplates: intended use, known limitations, data used in development, evaluation and bias-testing results, and the information you need for your notices. Vendors that cannot or will not supply this by 1 October are a decision point — mitigate, replace, or accept and document.

Week 4: notices. Draft the interaction notice and the written pre-decision notice. Place the interaction notice at the point the applicant or employee encounters the process, not in a policy document nobody opens. Confirm the pre-decision notice is delivered before the decision, which usually means a workflow change rather than a template.

Week 5: provenance, if applicable. If you are a covered provider over one million monthly users, document your technical feasibility assessment and implement marking for audio, image and video output. Align the implementation with your EU AI Act Article 50 work.

Week 6: the layoff determination process, and contracts. Assign ownership of the AI-causation determination and define the evidentiary basis. Brief legal, HR and finance. Separately, open the contract amendment cycle with AEDT vendors on indemnification, audit rights and bias-testing evidence — that cycle will not close in six weeks, but it should be open.

The wider pattern

Connecticut is now the third significant state framework governing automated employment decisions, alongside Colorado’s AI Act and New York City’s Local Law 144. The obligations are converging: notice before use, documentation flowing from developer to deployer, and the removal of the vendor as a liability shield.

That convergence points at a single design decision worth making once. Build the AEDT inventory, the vendor documentation pack, the notice templates and the bias-testing evidence file as a jurisdiction-neutral capability, then map it to Connecticut, Colorado, NYC and — for organisations with European operations — the EU AI Act’s Annex III employment provisions when they arrive on 2 December 2027.

Organisations that instead build a Connecticut project in September will be building a Colorado project in the winter and a New York project after that, four times over, with four sets of artefacts that describe the same systems.

The systems have not changed. Only the number of regulators asking about them has.

This article is provided for informational purposes only and does not constitute legal advice.