Does New Jersey have an AI hiring law like NYC Local Law 144?

Not in the same form, and that difference is often misread as an absence of risk. New Jersey's approach is that the Law Against Discrimination already applies to decisions made with automated tools, so no new statute was required to create liability. The practical consequence is the opposite of relief: Local Law 144 gives a covered employer a defined compliance path — an annual bias audit, a published summary, candidate notice — and completing it is visible and finite. The LAD gives you a liability standard with no safe harbour attached. There is no audit you can file that ends the inquiry, and the remedies available under the LAD are considerably broader than the penalty schedule of a notice ordinance.

Can we be liable if the discrimination was the vendor's model, not our decision?

Yes, on both sides of that relationship. An employer or housing provider that uses a tool remains responsible for the decisions it makes with that tool; outsourcing the analysis does not outsource the duty, and 'the model did it' has never been a defence to a disparate impact claim. Separately, the LAD contains an aiding and abetting provision that can reach a party who assists an unlawful practice, which is the theory under which a tool vendor may be pulled into a case rather than watching from the sidelines. The Mobley litigation against Workday made that exposure concrete nationally, and a vendor selling screening software into New Jersey should assume the question will be asked.

Does disparate impact under the LAD require proof of intent?

No. A facially neutral practice that produces a disproportionate adverse effect on a protected class can violate the LAD without anyone intending that result, which is exactly the shape of algorithmic harm. A model trained on historical outcomes reproduces historical patterns, and nobody in the room chose that. Intent is not the question; effect and justification are. That also frames the defence: whether the practice is job-related and consistent with business necessity, and whether a less discriminatory alternative was available. The second half of that test is why 'we never tested alternatives' is a worse answer than most employers realise.

What about disability — is that a separate problem from race and sex disparity?

It is a distinct failure mode and frequently the first one a real complaint raises. Statistical disparity analysis across large groups does not surface the individual whose disability caused a video-interview tool, a timed assessment, or a gamified screen to score them out. The LAD's disability provisions include a duty of reasonable accommodation, and an automated screen that offers no route to request one screens out the person before any human sees the request. Publishing a clear accommodation path at the point of the automated step — not buried in a careers FAQ — and training recruiters to honour it is the cheapest risk reduction available in this whole area.

New Jersey LAD and AI: Discrimination Liability Without a New AI Law

The "Existing Law Already Applies" Move

There are two ways a state can regulate algorithmic decisions. It can legislate a new regime with defined duties — impact assessments, notices, audits — as Colorado did. Or it can state, through its enforcement agency, that its long-standing civil rights statute already covers the conduct and proceed to enforce it. New Jersey took the second route, and the Division on Civil Rights has been explicit that the LAD reaches discrimination arising from automated decision-making tools in employment, housing, and places of public accommodation.

The second route produces a very different compliance experience. A new statute tells you what to do; you do it, and you have a record showing you did. An enforcement position under existing law tells you what you may not cause, and leaves the method to you. Our algorithmic hiring compliance overview maps the statutory regimes; this piece is about the states that decided they did not need one.

New Jersey is not alone in that posture, which is why this matters well beyond one state. The same reasoning is available to every attorney general sitting on a civil rights statute and a fair-housing law, and it is a far faster path than a legislative session.

Where the Exposure Actually Sits

Hiring and promotion screens. Resume parsers, ranking models, assessment scoring and video interview analysis. The disparity does not have to come from a protected attribute in the feature set; a proxy reproduces it perfectly well. See our note on the four-fifths rule and why passing it is not the end of the analysis.
Tenant screening. Housing is squarely within the LAD, and automated tenant screening combines the two features that attract claims: a scoring model built on records that carry historical disparity, and a decision communicated as a number.
Public accommodations. The LAD's reach here is broad, and chatbots, eligibility screens and automated service triage are decisions about who gets served — a category that most compliance programmes never inventory at all.
Advertising and audience selection. A model that decides who sees a job or housing listing has made a distribution decision with the same effect as a screening decision, which is the theory behind ad delivery claims.
Accommodation chokepoints. Any automated step with no visible route to ask a human for an adjustment. This is where disability claims begin, and it is invisible to aggregate fairness metrics.

Why "We Bought It, We Did Not Build It" Fails

The most common structure in the mid-market is an employer using a vendor's screening product with no visibility into its behaviour. Both parties tend to assume the other holds the risk. Under a disparate impact framework, both hold some.

The user of the tool made the decision. That is the beginning and end of the employer's or landlord's position, and it does not improve with the observation that a third party supplied the score. What the employer can do is demonstrate care: that it asked what the tool does, tested outcomes on its own applicant population, and acted on what it found.

The vendor's exposure runs through a different door. The LAD's aiding and abetting provision reaches a party who knowingly gives substantial assistance to an unlawful practice, and the national Mobley v. Workday litigation established that courts will entertain theories placing a screening vendor in the role of agent rather than bystander. A vendor whose contract disclaims all responsibility for outcomes while its marketing promises better candidate quality is holding two positions that do not sit together comfortably in front of a factfinder.

What to Do When There Is No Checklist

The absence of a prescribed compliance path is not an absence of good answers. The question a regulator or plaintiff will ask is what you knew about your tool's effects and what you did about it. Five artifacts answer that.

An inventory. Every automated step in hiring, tenancy, and service delivery, with the decision it influences and the weight it carries. Most companies cannot produce this, and it is the first document requested.
Outcome testing on your own population. Selection rates by protected class at each stage, not a vendor fairness score computed on someone else's data. Run it at the stage level: aggregate parity often hides a single stage doing all the damage.
A documented business-necessity rationale. Why this screen predicts performance in this role, in writing, before you need it. A validation study is better; a reasoned memo is far better than nothing.
An alternatives record. What less-discriminatory configurations you considered — a different threshold, fewer features, human review at the margin — and why you chose as you did. The less-discriminatory-alternative prong is where defences most often fail.
A visible accommodation route. At the automated step itself, with a named owner and a response time. Cheap, fast, and it resolves the complaint that is most likely to arrive first.

Run the testing under privilege where you can, and decide in advance what you will do if the numbers are bad — because discovering a disparity and leaving the tool running is materially worse than never having looked. That is the one genuine argument against testing, and the answer to it is a remediation plan, not ignorance.

Frequently Asked Questions

We are based in another state but hire remote workers in New Jersey. Does the LAD reach us?

Assume it can. Employment discrimination statutes generally follow the work and the applicant rather than the employer's headquarters, and remote hiring places candidates within a state's protective interest without any office there. This is the same analysis that caught out-of-state employers under NYC's rules, and the practical response is identical: apply your best process everywhere rather than maintaining a state-by-state matrix of screening behaviour. Configuring a hiring funnel to behave differently by candidate location is both operationally fragile and unattractive to explain.

Does the federal retreat from disparate impact theory reduce our New Jersey risk?

No, and this is a genuinely important point. Federal enforcement priorities are one input; state civil rights statutes are independent of them, are enforced by state agencies, and carry private rights of action that do not depend on any federal agency's posture. New Jersey's statute and its remedies are unchanged by shifts in federal emphasis. If anything the practical effect runs the other way, as state enforcement and private plaintiffs occupy space that federal agencies step back from.

Our vendor completed a bias audit for NYC. Does that cover New Jersey?

It is useful evidence and it is not coverage. A Local Law 144 audit is a specific, narrow exercise: impact ratios on defined categories, often computed on the vendor's aggregated data rather than your applicant pool, for the tool as configured generally rather than as you configured it. That is worth having in the file. It does not establish job-relatedness for your role, does not examine your stage-level outcomes, and says nothing about disability screen-out. Ask for the audit, then run your own numbers.

How far does the aiding and abetting theory really extend to software vendors?

It is unsettled and moving, which is itself the planning answer. The theories being tested nationally frame the vendor as performing a function traditionally performed by the employer, or as substantially assisting a practice it knows produces disparity. Whether a particular vendor falls inside depends heavily on how much discretion the tool exercises and how much the vendor knew. A vendor that configures the screen, sets thresholds, and reports on outcomes is a different defendant from one shipping a generic tool a customer tuned alone.

Should we just stop using automated screening in New Jersey?

Rarely the right call, and often a worse one. Manual screening is not free of bias, produces no record of its criteria, and is harder to test than a model. The advantage of an automated step is that it is measurable: you can compute selection rates by stage, change a threshold, and observe the effect. A screen you can audit and tune is a stronger position than a recruiter's judgment you cannot reconstruct. The risk is not automation; it is automation nobody has measured.

The Map Is Not the Territory

Compliance teams build state matrices with a column for "has an AI law." New Jersey's column is blank, and its exposure is not. Every state with a civil rights statute, a fair housing act and an attorney general able to read them the way New Jersey's did is a state where automated decisions already carry liability — no legislature required.

So build the matrix around the decisions rather than the statutes. If a tool influences who gets hired, housed or served, the question is whether you can show what it does to whom, everywhere you operate. That artifact is portable. A per-state checklist is not.