Who must comply with NYC Local Law 144?

Any employer or employment agency using an automated employment decision tool (AEDT) to screen candidates for a job located in New York City, regardless of where the employer is headquartered. They must run an independent bias audit within the prior year, publish the summary, and notify candidates at least 10 business days in advance.

When do EU AI Act obligations apply to hiring systems?

AI systems used for recruitment and employee selection are high-risk under Annex III 4(a). High-risk obligations for hiring systems apply from August 2, 2026, covering risk management, data governance, human oversight, transparency, logging, and post-market monitoring.

An Automated Employment Decision Tool is any computational process derived from machine learning, statistics, or AI that produces a score, classification, or recommendation used to substantially assist or replace discretionary hiring or promotion decisions.

What are the penalties for NYC Local Law 144 non-compliance?

Fines start at $500 for a first violation and reach $1,500 for each subsequent violation. Each day of non-compliant use and each missed candidate notice counts as a separate violation.

Can our AI vendor perform our bias audit?

No. NYC Local Law 144 requires an independent auditor with no role in building, selling, or using the tool and no financial interest in it. Vendors can supply data but cannot serve as the independent auditor.

Beyond the Checkbox: How to Fix Bias in AI Hiring Systems, Not Just Detect It

Summary: A bias audit tells you whether a hiring tool treats groups differently. It does not tell you why, or how to fix it. Remediation is a separate discipline. Real fixes come from understanding where bias enters the system: the data it learned from, the target it was trained to predict, the thresholds it applies, and the human decisions wrapped around it. Reducing bias durably means addressing those sources with evidence, not just re-running the audit and hoping the number improves.

Key takeaways

Detection and remediation are different skills. Passing an audit is not the same as building a fair system.
Bias usually enters through the data, the prediction target, the decision thresholds, or the humans around the tool.
Fixing the target the model optimizes for is often more powerful than adjusting the model itself.
Human oversight only reduces bias when it is designed to counter it, not when it merely sits on top.
Durable fairness comes from changing how the system defines a good hire, not from cosmetic tuning.

The gap between detecting and fixing

Most conversations about AI hiring bias stop at detection. A bias audit produces impact ratios, some below the 0.8 benchmark, and the company now knows it has a problem. Then comes the harder question that the audit does not answer: what do we actually change?

This is where many programs stall. Detection is a measurement problem with an established method. Remediation is a design problem with no single formula. It requires understanding the mechanism behind the disparity, and mechanisms differ from tool to tool. Treating the audit as the finish line leaves the underlying problem in place and sets up a repeat finding next year.

Where bias actually enters a hiring system

Bias in AI hiring is rarely a single villainous line of code. It enters through four main doors.

The data it learned from

If a model is trained on a company's past hiring decisions, and those decisions favored one group, the model learns to reproduce that pattern. It is not malfunctioning. It is doing exactly what it was asked, which is to predict the historical outcome. Historical data encodes historical bias, and a faithful model inherits it.

The target it was trained to predict

This is the most underappreciated source. A model does not predict good employee in the abstract. It predicts a proxy that someone chose, such as who got hired, who received a high performance rating, or who stayed two years. If that proxy is itself biased, the model optimizes toward bias with perfect fidelity. Changing the target, choosing a fairer, more job-relevant outcome to predict, is often more powerful than any adjustment to the model.

The thresholds it applies

Where you set the cutoff for advancing a candidate changes who makes it through. A threshold that looks neutral can produce very different pass rates across groups depending on the shape of each group's score distribution. Threshold design is a lever that is frequently overlooked because it feels like a technicality rather than a fairness decision.

The humans wrapped around it

A tool sits inside a human process. Recruiters decide how much to trust the score, which candidates to override, and how to interpret ambiguous cases. If those human decisions are themselves patterned by bias, they can amplify or dampen whatever the tool does. Fixing the model without addressing the human loop leaves half the system untouched.

Why human oversight is not automatically a fix

The laws require human oversight, and companies often assume that a person in the loop neutralizes bias. It does not, by default. Oversight reduces bias only when it is designed to counter specific failure modes. A reviewer who simply confirms the tool's ranking adds a rubber stamp, not a safeguard. Worse, automation bias, the tendency to over-trust machine output, can make human review less critical over time, not more.

Effective oversight is structured. It defines what the reviewer is looking for, gives them the information to challenge the tool, and measures whether their interventions actually change outcomes. Oversight that is designed rather than assumed is what turns a legal requirement into a genuine control.

A behavioral approach to remediation

Fairer hiring is partly a technical problem and partly a behavioral one, because a hiring system is a chain of human and machine decisions. A behavioral approach asks not only whether the model is fair in isolation, but whether the whole decision environment produces fair outcomes.

That means examining the choice architecture around the tool. How results are presented to recruiters shapes how they act on them. Whether scores are shown as precise numbers or ranges changes how much authority they carry. Whether reviewers see demographic information at the wrong moment can introduce bias that the model itself avoided. Small design choices in how information flows to decision makers can move outcomes as much as retraining the model. This is the layer that purely technical remediation misses, and it is where behavioral science earns its place in the work.

What durable remediation looks like

A remediation that lasts changes how the system defines and pursues a good hire, rather than nudging a number just past a threshold. In practice that means re-examining the prediction target for job-relevance and fairness, auditing the training data for inherited patterns, redesigning thresholds with their group-level effects in mind, and structuring human oversight to actually catch what the model misses. It ends with a re-audit that confirms the disparity closed for a reason you can explain, not by coincidence.

The difference between cosmetic and durable remediation shows up the following year. A cosmetic fix drifts back. A durable fix holds because the mechanism that produced the bias was addressed.

Frequently asked questions

Isn't passing the bias audit enough? Passing the audit means the current numbers clear the benchmark. It does not mean the underlying mechanism is fixed, and unaddressed mechanisms tend to resurface.

What is the single most effective fix? Often it is changing the target the model is trained to predict, because a biased target guarantees biased output no matter how good the model is.

Does adding a human reviewer solve bias? Only if the oversight is designed to counter specific failure modes. Unstructured review can add a rubber stamp and invite automation bias.

Why involve behavioral science? Because a hiring system is a chain of human decisions around a tool. How results are presented and acted on affects outcomes as much as the model itself.

How do we know a fix worked? A re-audit should show the disparity closed for a reason you can explain, and it should hold in the following year's audit rather than drifting back.

Detecting bias is the easy half. Fixing it durably is where hiring actually improves, and where compliance becomes an asset rather than a chore. If your audit surfaced disparities and you want a remediation grounded in how these systems really produce bias, that is the work we specialize in.