Connecticut SB 2 AI Controls Mapping: Public Act 26-15 and the Amended CTDPA at the Request Boundary

Connecticut's AI obligations sit in two statutes, and neither is the one most compliance plans were written against. Substitute Senate Bill 2 passed the Senate on 14 May 2025 and died in the House for the second session running. Governor Ned Lamont signed Substitute Senate Bill 5 as Public Act 26-15 on 2 June 2026, and sweeping amendments to the Connecticut Data Privacy Act took effect on 1 July 2026.

Connecticut compliance work tends to concentrate on assessments and disclosures, both of which describe a system before it runs. The mapping from an obligation to the control that enforces it, at the point personal data actually leaves the network, usually goes unwritten. I want to set that out obligation by obligation, because the profiling impact assessment duty attached to processing created on or after 1 August 2026 and the Public Act 26-15 provisions begin landing on 1 October 2026.

The control point: one HTTPS request

Every mapping below resolves to the same moment: an HTTPS POST carrying a context window to a model endpoint. That request is where identity is known, where classification is evaluable, where a consumer's opt-out state can be checked, and where a decision is recorded or lost. A control described in an impact assessment names the obligation without enforcing it.

Profiling for legally significant decisions (amended CTDPA)

The amended CTDPA requires controllers profiling consumers for legally significant decisions to run a dedicated impact assessment, separate from the data protection assessment, for processing created or generated on or after 1 August 2026. Carrying the operational half of that duty takes a per-decision record binding the input prompt to the output and the applied policy, generated by enforcement rather than by a questionnaire. The evidence is decision lineage a reviewer can query for a named consumer on a named date.

The opt-out, with "solely" removed

The 2026 amendments struck "solely" from the automated decision-making provision, extending the opt-out to profiling in furtherance of any decision producing a legal or similarly significant effect, including decisions with a human involved. A human reviewer no longer takes the processing out of scope.

The control is suppression state evaluated in the request path: before a prompt is assembled, the caller's policy context carries whether this consumer has opted out, and the request is routed, redacted, or refused accordingly. The evidence is a per-request record showing the flag was checked and what followed. Enforcing this at the decision desk instead of the request boundary leaves the prompt already sent.

Employment decisions (Public Act 26-15)

Public Act 26-15 reaches employment decision-making among its covered uses, turning on whether an AI system was a substantial factor in a consequential decision. The control is identity binding plus route tagging, so requests originating from a hiring or promotion workflow are marked as such at the moment they are made. The evidence is a per-transaction link from the named person or agent to the AI call that informed the outcome, which is the post-authentication gap closed at the boundary.

Generative content provenance

The act's content provenance obligations attach to output rather than to personal data, so the enforcement point moves from request to response. Enforcement becomes a marking and recording step at the response boundary, applied before content leaves the system. The evidence is a record of what was generated, by which model, for which caller, and whether the marking was applied, which is the same pattern California SB 942 controls mapping describes for the covered-provider disclosure duties.

AI companions and platforms used by minors

Public Act 26-15 addresses AI companions and online platforms used by minors, with obligations phasing in through 2027 and January 2028. The control is age and context-aware policy on the model channel, so a route serving a minor carries a different permitted set than a route serving an adult account. The evidence is a per-route policy record joined to the traffic that ran under it.

Frontier model obligations

The act also covers frontier AI models, which affects developers rather than most deployers. For an enterprise consuming frontier models through an API, the practical control is the same destination-aware routing every other row needs: which endpoint received the request, under which policy, carrying which classification. The evidence is a destination and classification log, produced by AI data residency controls in the request path.

The mapping at a glance

Enforcement sits with the Attorney General

The CTDPA is enforced by the Connecticut Attorney General with no private right of action, which shapes what evidence has to look like. Enforcement arrives as an inquiry in writing, and the answer is a document produced from records rather than a description of a control design. My frank opinion: the removal of "solely" is the single most under-instrumented change in the 2026 amendments, because it converts an exemption that a workflow diagram used to satisfy into a runtime check that most request paths cannot perform today.

DeepInspect

Each row above resolves to an enforcement point, and DeepInspect is that point. It sits inline between your users or agents and the LLM APIs they call, evaluates identity, classification, consumer suppression state, and destination on every request, enforces per-role and per-route policy, and writes a signed record of the decision before the response returns. The Connecticut mapping stops being a set of filed assessments and becomes a running control with evidence attached.

For a controller facing Attorney General scrutiny of profiling decisions and opt-out handling, that means each obligation has a per-decision record standing behind it. The Connecticut AI audit evidence view covers what each artifact looks like under inquiry. Book a technical deep dive at deepinspect.ai.