New Jersey Employers Face an AI Hiring Notice Patchwork

Consider requisition Req-427, a hybrid data-analyst job in Manhattan run by a New Jersey employer. Its hiring system parses resumes, administers a skills test, assigns a score, and moves the highest-ranked applicants to a recorded video interview. A recruiter can override the ranking, although the score normally determines who advances.

That workflow may trigger New York City’s automated employment decision tool law. If the same employer copies Req-427 for an Illinois-based position, the video stage may invoke Illinois consent and deletion requirements. A Colorado version could eventually require an impact assessment and a process for correcting data and appealing an adverse decision. New Jersey headquarters do not settle any of those questions.

The practical unit of compliance is therefore not the company or software contract. It is each requisition, including where the job is based, where the applicant resides, which automated output affects the decision, and what happens after the system rejects someone.

This explainer maps that workflow from public statutes, regulations, and agency materials. It is not legal advice, and employers should verify effective dates and geographic coverage before relying on it.

Start with the output, not the AI label

An automated employment decision tool, or AEDT, is software that produces a score, classification, or recommendation used to make an employment decision. The exact statutory definition varies. A chatbot that answers questions may sit outside one law, while a resume ranker that determines who receives an interview may fall squarely inside it.

For Req-427, document the sequence before reviewing geography: the parser extracts education and work history; the test produces a score; the ranking combines those fields; the recruiter sees a sorted queue; rejected applicants receive an automated email. Record whether recruiters routinely review lower-ranked applicants, whether an override is possible, and whether the video module analyzes facial, vocal, or word-choice features rather than merely recording the interview.

Those details decide whether a law reaches the tool. New York City’s rules cover a qualifying computational process only when its output “substantially assist[s] or replace[s] discretionary decision making.” The city’s rules treat an output as substantially assisting when an employer relies on it alone, weighs it more heavily than other criteria, or uses it to overrule conclusions drawn from other factors.

A vendor calling its product “decision support” does not answer that test. Neither does the presence of an override button if recruiters rarely use it and the ranked queue controls who advances.

The decision tree for each requisition

Run every posting through the following branches before opening applications:

1. Identify a consequential output. If software only schedules interviews or stores documents, record that conclusion and the supporting facts. If it ranks, recommends, scores, or filters people, continue.

1. Map the job and applicant geography. Flag a New York City work location, an applicant residing in the city, an Illinois-based position using an AI-analyzed video interview, a Maryland interview using facial recognition, and a Colorado employment decision. Remote jobs need an explicit location rule rather than a blank field.

1. Attach the pre-use requirement. This may be a published bias-audit summary, advance notice, an explanation of analyzed characteristics, or written consent. “Pre-use” must be implemented as a system gate, not left in a recruiter handbook.

1. Attach the response route. Record where applicants request accommodation, an alternative process, deletion, data correction, or human review; set an owner and deadline; then test whether the rejection email points to that route.

1. Preserve the evidence. Keep the tool version, configuration, notice text, delivery timestamp, applicant response, score, recruiter action, and override. Without those records, an employer may know its policy but be unable to show what happened in Req-427.

This tree should sit in the applicant-tracking system’s requisition setup. A spreadsheet reviewed after candidates have been scored cannot deliver notice before use or stop an unapproved video-analysis module.

New

York City adds the clearest audit-and-notice gate

New York City Local Law 144 bars covered employers and employment agencies from using an AEDT unless it has undergone a bias audit no more than one year before use and a summary of the results is publicly available. The audit must be performed by an independent auditor under the city’s rules.

The audit is a statistical assessment, not a general promise of fairness. Depending on the tool’s output, the rules call for selection or scoring rates and impact ratios across sex, race or ethnicity, and intersectional categories. Historical data from the employer should be used when sufficient data exists; test data can enter under circumstances specified by the rules, with that choice disclosed.

For a covered candidate residing in New York City, the law requires notice at least 10 business days before use. The notice must say that an AEDT will be used and identify “the job qualifications and characteristics” it will use in the assessment. It must also allow the candidate to request an alternative selection process or accommodation.

Local Law 144 does not itself require the employer to grant every requested alternative. Disability law may independently require a reasonable accommodation, which is why the request cannot disappear into a generic recruiting inbox.

Req-427 needs a release gate: no resume score enters the recruiter’s queue until the employer has matched the deployed configuration to a current audit, posted the required summary, and delivered the candidate notice on time. Reusing a vendor’s audit without checking the employer’s configuration, data, and use can leave a gap between the assessed tool and the deployed one.

Video interviews create a different branch

Illinois regulates a narrower workflow under its Artificial Intelligence Video Interview Act. An employer asking an applicant for an Illinois-based position to record a video interview, then using AI to analyze it, must notify the applicant that AI may analyze the video, explain how the system works and what general characteristics it evaluates, and obtain consent before the interview. The employer may not use the AI analysis without that consent.

The law also limits sharing. If an applicant asks for deletion, the employer must delete the interview and instruct others holding copies or backups to delete them within the statutory period. A separate reporting provision applies when an employer relies solely on AI analysis to decide which applicants receive an in-person interview.

Maryland takes another narrow approach: an employer may not use a facial-recognition service to create a facial template during an applicant interview unless the applicant signs a consent waiver. That is not an annual bias audit, and it should not be represented as one.

For Req-427, the operational choice is clean. The employer can disable automated video analysis for Illinois and Maryland interviews while retaining ordinary recording, subject to other applicable rules, or it can build jurisdiction-specific consent and deletion controls. Buying one national video workflow is cheaper to administer, but only if its strictest controls are acceptable everywhere.

Colorado adds impact assessment and contestability

Colorado’s enacted AI Act covers certain high-risk AI systems used in consequential employment decisions. Its commencement has been delayed, so employers should verify the operative date and implementing rules; requirements that are scheduled but not yet effective should not be described internally as present enforcement.

The enacted framework is broader than a video-interview consent form. Deployers must use risk-management policies, complete impact assessments at required intervals and after specified modifications, and notify consumers when a high-risk system will make or substantially factor into a consequential decision. Following an adverse decision, the framework calls for a statement of principal reasons, disclosure about the data involved, an opportunity to correct inaccurate personal data, and an appeal that allows human review when technically feasible.

That makes the fallback part of system design. If the employer cannot reconstruct which inputs drove Req-427’s rejection, a human reviewer receives little more than the original score. Logging every model feature may expose sensitive data and vendor intellectual property, but logging nothing leaves the employer unable to explain or contest the result. The workable middle is a decision record tied to the candidate, tool version, material input categories, output, threshold, and recruiter action.

New Jersey still supplies the baseline

New Jersey has considered legislation aimed specifically at automated employment decisions, including notice and bias-review concepts, but a proposal is not an enacted compliance duty. Employers should monitor bills without placing proposed requirements in the same column as New York City’s enforceable law or Illinois’s existing video-interview statute.

The absence of a comprehensive New Jersey AEDT law does not remove discrimination risk. The New Jersey Law Against Discrimination and federal employment laws apply to hiring outcomes regardless of whether a person, model, or vendor produced the ranking. The federal Americans with Disabilities Act can also require reasonable accommodation when an assessment disadvantages an applicant because of a disability.

A bias audit is evidence, not immunity. Group-level impact ratios may reveal that one cohort advances at a lower rate, but they do not establish that every tested feature is job-related, detect every disability barrier, or prove that the deployed configuration matches the audited version.

For the New Jersey employer operating Req-427, the durable setup is a requisition-level control plane: geography fields that cannot be skipped, a registry of decision-making tools, version-linked audits, timed notice delivery, a consent gate for video analysis, and a case queue for accommodation, deletion, correction, and appeal requests. That costs recruiter time and engineering work. It is cheaper than discovering after rejection that the system never recorded which rule applied.

Questions people ask

Does

New York City’s law apply because the employer is in New Jersey?

Not by headquarters alone. Coverage turns on the tool’s use and its connection to New York City employment and covered candidates, with agency rules and guidance supplying important geographic details. A New Jersey employer filling a Manhattan role should screen the requisition before use rather than assume crossing the Hudson removes the obligation.

Is a vendor’s bias audit enough for every employer using its tool?

Not automatically. The employer must determine whether the audit covers the tool version, configuration, output, and use that it deploys, and New York City requires a current audit plus a public summary before use. Vendor documentation can support that review, but a generic fairness report is not necessarily the required audit.

Must an employer offer a human alternative to every AI assessment?

No single rule creates that universal duty. New York City requires a route for requesting an alternative process or accommodation but does not itself require every alternative to be granted; disability law may require accommodation. Colorado’s enacted framework calls for an appeal with human review when technically feasible once its relevant provisions become operative.