AI résumé screening, bias, and New York City's AEDT law

This page describes the rules in general terms and is not legal advice; whether a specific tool is in scope is a question for your counsel.

What Local Law 144 covers

Local Law 144 of 2021 applies to an automated employment decision tool: broadly, a computational process derived from machine learning, statistical modeling, data analytics or artificial intelligence that produces a simplified output — a score, a classification or a recommendation — used to substantially assist or replace discretionary decision making for hiring or promotion. The Department of Consumer and Worker Protection adopted implementing rules in 2023, began enforcing the law on July 5, 2023, and enforces it for positions in New York City.

Two obligations follow. The tool must have had an independent bias audit within the past year, and a summary of that audit's results must be publicly available. Candidates and employees who live in the city must be notified at least ten business days before the tool is used on them, and told what qualifications and characteristics it assesses. The audit itself compares selection or scoring rates across sex and race or ethnicity categories, including the intersections of the two.

Where bias enters résumé screening

The obvious mechanism — a model trained on who was hired before — is not the only one. A ranking function built on features that correlate with protected characteristics does the same work through proxies: school, ZIP code, employment gaps, name and language patterns, the prestige of previous employers. Ranking compounds it. A model that is only slightly skewed still concentrates that skew at the top of a list, and a reviewer who reads the first ten of two hundred never sees the effect.

Scoring a candidate versus showing a reviewer evidence

The statutory hinge is the simplified output and whether it substantially assists the decision. A system that retrieves what public sources say about a claim, shows the reviewer the sources, and produces no score, rank or recommendation about the person is built to sit on the other side of that line — but that is a design intent, not a determination, and the determination depends on how your team actually uses the output. If reviewers treat a flag as a verdict, the label on the software will not matter much.

Gram is built this way deliberately: it surfaces the claim, the evidence and the supported wording, and leaves the judgment with the reviewer.

What to ask a vendor

Ask whether the vendor considers the tool an AEDT under Local Law 144, and why. Ask for the published bias audit, its date, and who the independent auditor was. Ask what the tool outputs for each candidate, whether anything is ranked or scored, and what data it was trained or tuned on. Ask to see the evidence behind any individual output, and what happens when the evidence is absent. Ask what notice text they provide and who is responsible for sending it.

Federal guidance has shifted; the law has not

The EEOC's May 2023 technical assistance on artificial intelligence under Title VII was removed from its website in 2025, along with other AI-related material. That withdrawal changed the guidance, not the statute: a selection procedure that disproportionately excludes people on a protected basis still has to be job-related and consistent with business necessity, as the EEOC's standing guidance on employment tests and selection procedures explains. State and local rules in this area are still being written, so check the jurisdictions your candidates are in rather than assuming the federal position settles it.