Bias in Hiring AI: The Bias Audit and NYC Local Law 144, Explained

Bias in hiring AI is both a real risk to candidates and a real legal exposure for employers. When an automated tool helps decide who advances in a hiring process, it can quietly carry forward patterns from the data it learned on, and in New York City that risk now comes with a specific compliance regime attached. This guide explains, in plain terms, what NYC Local Law 144 requires, how it connects to longstanding EEOC expectations, what a bias audit actually is, and why structured, rubric-based AI interviews with a human in the loop are a sound way to reduce both the bias and the legal risk. This is an explainer, not legal advice. For your own situation, talk to qualified employment counsel.

Why AI can introduce hiring bias

An automated hiring tool learns patterns from historical data, and historical hiring data reflects historical decisions, including biased ones. If a model is trained to mimic past hires, it can reproduce who was favored before, even when no protected characteristic is used as an input directly. Bias can also enter through proxies, where something seemingly neutral like a zip code, a school, or a speech pattern correlates with race, gender, age, or disability. The harm is not hypothetical. A tool that scores candidates can systematically disadvantage a protected group at scale, faster and more uniformly than any single biased manager ever could. That scale is exactly why regulators and the EEOC have focused on it.

What NYC Local Law 144 actually requires

NYC Local Law 144, in effect since 2023, regulates the use of what it calls an automated employment decision tool, or AEDT, when used to screen candidates for employment or promotion in New York City. The law sets out a few core obligations for employers and employment agencies that use a covered tool.

• A bias audit. Before using a covered AEDT, and within one year of the audit on an ongoing basis, the employer must have the tool independently bias-audited. The audit calculates selection or scoring rates across sex categories and across race and ethnicity categories, and reports impact ratios that show whether any group is selected at a meaningfully lower rate than the most-selected group.
• Public posting. A summary of the most recent bias audit results, and the distribution date of the tool, must be published on the employer's website where candidates can find it.
• Candidate notice. Candidates who live in NYC must be notified at least ten business days before the tool is used that an AEDT will be part of the process, told which job qualifications and characteristics it assesses, and given information on how to request an alternative selection process or accommodation.

A few things commonly trip people up. The bias audit must be performed by an independent auditor, not the vendor grading its own homework. The audit looks at outcomes across groups, not at the model's internals. And the law is specifically about tools that substantially assist or replace a human in the selection decision, which is one reason keeping a genuine human decision-maker in the loop matters so much.

How this connects to the EEOC

Local Law 144 is a local rule, but it sits on top of much older and broader federal law. Title VII of the Civil Rights Act prohibits employment discrimination, and the EEOC has made clear that using an algorithmic or AI tool does not exempt an employer from those obligations. Under the longstanding disparate impact framework, a selection procedure that screens out a protected group at a substantially lower rate can be unlawful even with no intent to discriminate, unless the procedure is job-related and consistent with business necessity. The familiar four-fifths rule, where a selection rate for any group below 80 percent of the highest group's rate is a flag for adverse impact, comes from EEOC guidance and is the same logic the NYC impact ratios are built on. The practical takeaway is that an employer is responsible for the outcomes of its hiring tools, whoever built them.

What a bias audit looks like in practice

A bias audit is fundamentally an outcomes analysis. The auditor takes the tool's real selection or scoring data, groups candidates by sex and by race and ethnicity, and compares how often each group is advanced or scored highly. From those numbers come impact ratios, which surface whether the tool is favoring one group over another. A clean audit is not a one-time certificate. Hiring data shifts, models get retrained, and roles change, so the audit has to be repeated, which is why the law ties it to an annual cadence. The audit does not prove a tool is perfectly fair forever. It gives employers and candidates a documented, regular check on whether the tool's outcomes are skewed, and a paper trail if anyone asks.

A bias audit does not measure intentions. It measures outcomes across groups, which is the only thing that protects candidates and the only thing a regulator can verify.

Why structure and a human in the loop reduce the risk

Two design choices do more than anything else to lower bias risk. The first is structure. A structured interview asks every candidate the same job-related questions and scores them against the same defined rubric, which both reduces the subjective variance where bias hides and produces clean, comparable data that an audit can actually evaluate. Free-form interviews, by contrast, are hard to audit and notoriously easy for bias to slip into. The second choice is keeping a human in the loop. When a person reviews the evidence and makes the final call, the tool is informing a decision rather than making it, which is both better practice and central to how these rules are framed.

How InterviewAgent.ai is built for this

InterviewAgent.ai is designed around these principles. It conducts first-round screens using the same role-tailored structured questions for every candidate and scores each answer against a defined rubric, so the process is consistent and the resulting data is auditable. Candidates are clearly told that they are interviewing with an AI interviewer and consent before they begin, which supports the disclosure and notice expectations at the center of Local Law 144. Every interview is transcribed, so there is a record of exactly what was asked and answered. Most importantly, the agent produces a ranked shortlist with scores and quotes, but it never auto-hires or auto-rejects anyone. A human recruiter or hiring manager always makes the final advancement decision.

This human-in-the-loop, structured, transparent approach is the same posture that compliance teams favor, and it is why we frame the product around EEOC principles and NYC Local Law 144 rather than around the tool deciding for you. You can read more about how the screening and scoring work on our pages for structured interview software and candidate scoring.

None of this removes your own legal obligations. If you use any automated hiring tool in NYC or anywhere subject to similar rules, you should confirm your bias-audit, notice, and posting duties with qualified employment counsel, and make sure your end-to-end process, not just the tool, holds up. The good news is that the practices that keep you compliant, structure, transparency, regular auditing, and a real human decision, are the same ones that make your hiring fairer and better.