Connecticut Requires AI Notices, Not A Bias Audit

Connecticut’s enacted AI-employment provisions require disclosure, not an independent bias audit: a covered pre-decision notice has eight principal information fields, while the law creates no separate audit, impact-assessment filing, public audit posting, or testing cadence. The principal applicant and employee duties begin October 1, 2027, not July 1, 2026. A narrower AI-related WARN disclosure begins October 1, 2026. The controlling source is Public Act 26-15, enacted from Senate Bill 5; the supplied source record does not support identifying these requirements as SB 2.

That does not make implementation costless. Employers must identify covered tools, determine how outputs affect decisions, obtain configuration-specific information from vendors, and deliver accurate notices before covered decisions. But those are classification and notice-management costs—not a Connecticut statistical fairness-audit mandate.

Why The Broader Compliance View Sounds Plausible

The received view is that Connecticut joined a growing group of jurisdictions imposing a substantial new compliance program on employers using automated hiring systems. That view is right about the law’s reach. The framework can apply beyond recruitment to promotion, discipline, discharge, employment renewal, training or apprenticeship selection, tenure, and material employment terms or conditions.

It is also right that a notice cannot always be drafted from a vendor brochure. An employer may need to map personal-data sources, outputs, configured thresholds, recruiter behavior, and the role of human review. The written notice must explain how data is assessed and identify the technology’s trade name, among other fields. Legal analyses accordingly describe separate interaction and decision-notice duties within Connecticut’s employment-AI framework beginning in October 2027.

The overstatement is treating that operational work as though Connecticut imported New York City’s bias-audit model. Connecticut’s AEDT provisions do not independently require an audit, recurring statistical test, public audit summary, or impact-assessment filing. Bias testing may still be prudent, may matter in discrimination proceedings, and may be required elsewhere. It is not one of Connecticut’s disclosure duties.

Choose how your tool is used; the result identifies the Connecticut notices triggered by that workflow and separates them from audit obligations.

Connecticut Has Three Disclosure Tracks

The statute creates two duties involving automated employment-related decision technology, or AEDT, and a third duty tied to an otherwise-required Connecticut WARN filing.

The first two duties address different conduct. A back-end résumé ranker may trigger the pre-decision notice without communicating with an applicant. A covered chatbot may require an interaction disclosure but not a decision notice if it does not evaluate or materially influence a decision.

The obviousness exception applies only to the interaction issue. If a reasonable person would recognize that the interaction is automated, a separate interaction statement is generally unnecessary. It does not excuse a pre-decision notice when the system’s output substantially affects who advances.

The WARN provision is narrower still. It does not create an AI notice for every termination or reduction in force. It adds information when the employer already must make the relevant WARN filing. The separate October 1, 2026 and October 1, 2027 dates are also reflected in employment-law analysis of the phased requirements.

Coverage Depends On Function, Not An AI Label

Connecticut regulates covered uses of AEDT, not every product marketed as “AI,” “smart,” or “automated.” Reported definitions focus on technology that processes personal data, computationally produces an output, and uses that output as a substantial factor in—or to materially influence—an employment-related decision.

An output can include a prediction, recommendation, classification, ranking, score, constraint, or similar information. A substantial factor is described as one that meaningfully alters an individual’s outcome, although available commentary does not fully resolve the boundaries of “materially influences” in the AEDT definition.

A practical coverage review asks four questions:

1. Does the system process personal data about an applicant or employee?
2. Does it computationally produce a score, rank, classification, recommendation, prediction, constraint, or comparable output?
3. Is that output used in hiring or another covered employment decision?
4. Does the output meaningfully change eligibility, visibility, review order, advancement, or the eventual result?

Keeping a person in the loop does not automatically end the analysis. A recruiter may make the nominal final decision while routinely accepting a system’s shortlist. A ranking can materially affect outcomes by controlling who receives timely review even if it never automatically rejects anyone.

Independent human review can matter factually. If reviewers examine all relevant materials, can disregard the output, and do not allow it to meaningfully affect outcomes, coverage may be less likely. The supplied evidence establishes no bright-line human-review exemption.

Jurisdictional questions also remain. The source material does not resolve every case involving remote applicants, out-of-state employers, cross-border workflows, employer-size thresholds, or decisions made outside Connecticut. Those facts require review against the enacted language and any later official guidance.

Tool Categories Are Only Starting Points

A product can be covered in one configuration and outside the AEDT definition in another.

Routine word processing, spellcheck, navigation, web hosting, antivirus, anti-malware, spam filtering, and purely descriptive or statistical functions are reported as excluded when they do not influence an employment decision. The same data can present a different case if it is later fed into a model used for discipline, promotion, discharge, or another covered judgment.

A chatbot requires two separate checks. First, determine whether its direct interaction falls within the interaction provision and whether automation would be obvious to a reasonable person. Second, determine whether it scores answers, applies knockout rules, classifies eligibility, or recommends advancement. If it does, both notices may be implicated.

A direct interaction statement can be simple:

You are interacting with an automated recruiting assistant.

The evidence does not establish a mandatory screen location, delivery channel, translation rule, accessibility standard, or proof-of-receipt process. Making the statement noticeable at or before the interaction and retaining the displayed version are sensible controls, but they should not be represented as confirmed statutory specifications.

The Pre-Decision Notice Has Eight Principal Fields

When AEDT makes or is a substantial factor in a covered decision, the affected applicant or employee must receive written notice before the decision is made. The evidence supplies no fixed number of advance days.

The notice should contain these eight principal fields, as summarized in analysis of Connecticut’s covered pre-decision disclosures:

1. A statement that AEDT has been deployed.
2. The technology’s purpose in the workflow.
3. The nature or type of employment decision.
4. The AEDT’s trade name.
5. The categories of personal data processed.
6. The sources of that personal data.
7. How the data will be assessed in reaching the decision.
8. Contact information for the employer or other deployer.

Data categories and sources are different fields. “Résumé history and application answers” describes categories; “the applicant and an assessment provider” describes sources. The assessment explanation should state what the system does with the inputs, what output it produces, and how decision-makers use that output.

A workable notice structure is:

Notice of automated employment-related decision technology

[Employer name] has deployed automated employment-related decision technology for [decision type or stage]. Its purpose is to [specific purpose], and its trade name is [trade name].

The technology processes [personal-data categories] obtained from [data sources]. It assesses that information by [explain processing and output], and the resulting [score, rank, classification, recommendation, or other output] is used to [explain its role in the decision].

Questions may be directed to [monitored employer contact].

This is a drafting structure, not an official Connecticut form. It must be matched to the configured technology and actual workflow.

Qualifying trade secrets or information protected by state or federal law may reportedly be withheld, but the affected person must be told that information was withheld and given the legal basis. A vendor’s confidentiality preference or nondisclosure clause does not by itself establish trade-secret protection.

The evidence does not settle whether one notice can cover multiple tools, whether the interaction and pre-decision statements can always be combined, whether an evergreen notice is sufficient, or whether notice must be reissued after a configuration change.

Connecticut Notice Compliance Is Not Bias-Audit Compliance

The AEDT provisions do not establish an independent requirement to conduct a bias audit or impact assessment. They also do not clearly create a right to opt out, appeal, demand an individualized explanation, insist on a human decision-maker, or prohibit the tool’s use after notice.

That limit should not be read as an argument against testing. Disclosure does not immunize a discriminatory decision, and using AEDT is not a defense to an employment-discrimination complaint. Anti-bias testing and the employer’s response to results may be considered in a discrimination proceeding.

The accurate division is therefore:

• Connecticut notice compliance: classify the deployment, prepare interaction or pre-decision disclosures, supply the required fields, deliver before the decision, and coordinate with vendors.
• Decision-risk management: assess job relevance, validation, bias, accommodations, human oversight, escalation, and unexpected outcomes.

An employer may need both workstreams. Only the first is imposed by these Connecticut disclosure provisions. Separate jurisdictions or other laws can independently change that answer.

Vendors Must Supply Configuration-Specific Information

The framework distinguishes the deployer that puts technology into use from the developer that creates or intentionally and substantially modifies it. An employer using a third-party hiring platform will commonly be the deployer; the provider will commonly be the developer.

Developers generally must provide information needed for the disclosures when technology is marketed, configured, contracted for, sold, or licensed for materially influencing employment decisions. A developer may also expressly assume specified disclosure duties by contract, according to analysis of Public Act 26-15’s developer-deployer rules.

A promise to “support compliance” is too vague to allocate notice work. The contract should identify who supplies each required field, drafts and approves language, delivers the notice, retains delivery records, reports configuration changes, and corrects inaccurate information.

Employers should verify vendor statements against actual use. A platform sold as a recruiting assistant could merely schedule interviews for one customer while classifying applicants for another. Generic marketing material cannot establish the inputs, outputs, or decision influence of a configured deployment.

Changes to inputs, data sources, outputs, scoring logic, trade name, interface, or intended use should reopen the classification and notice review. That is a risk-management control, not a separately identified statutory command.

Enforcement Does Not Turn The Notice Into A Safe Harbor

Violations of the AEDT disclosure provisions are reported as unfair or deceptive trade practices enforceable exclusively by the Connecticut attorney general. The cited analyses report no private right of action for those particular notice violations.

For qualifying or curable violations occurring on or before December 31, 2027, legal summaries report a possible 60-day cure opportunity. Sources conflict over whether that opportunity is mandatory or discretionary, so an employer should not assume it will receive one. The reported enforcement structure and cure period are outlined in Connecticut AEDT enforcement analysis.

The lack of a private action under these notice provisions is narrow. Discrimination, privacy, disability, accommodation, contractual, and other legal duties may still apply. Notice compliance does not validate the underlying decision.

The Practical To-Do List Starts With Inventory

Employers should inventory tools used in recruiting, performance management, learning, promotion, discipline, renewal, and discharge. For each deployment, record the product and trade name, affected population, personal data, data sources, outputs, decision stage, degree of influence, automated thresholds, direct interactions, and human-review process.

Classify each use as requiring an interaction disclosure, a pre-decision notice, both, neither, or legal review. Document the function rather than relying on labels such as “scheduling,” “analytics,” or “productivity.”

For potentially covered decision tools, obtain all eight notice fields and assign a delivery owner. A back-end ranker may need a notice embedded in the application or screening workflow even though applicants never interact with the ranking system.

Maintain practical records of the approved notice, enabled features, delivery point, vendor information, relevant changes, and responsible owner. The supplied evidence establishes no specific statutory retention period, so retention should be set under applicable employment, privacy, litigation-hold, and records-management requirements.

Finally, keep the WARN checkpoint separate. Beginning October 1, 2026, an employer already making the relevant Connecticut WARN filing must determine whether the workforce reduction relates to AI use or another technological change. The principal interaction and pre-decision duties follow on October 1, 2027, a phased schedule also described in multi-stage compliance coverage.

The applicant and employee provisions remain future-facing and may be affected by amendments or official guidance before they become operative. Templates and workflows should be checked against the enacted act and current Connecticut guidance before use.

This material is informational, not legal, HR, or employment advice, consistent with HRaizon’s Terms & Conditions. Application depends on the employer, technology, workflow, person, and location.