NYC Hiring-Bot Notices Do Not Replace Bias Audits

Consider a resume-ranking system used for a customer-support job in Manhattan. It reads applications, assigns each applicant a score, and sends the highest-ranked group to a recruiter. Before that score enters the hiring decision, the employer has two operational jobs: establish that an applicable independent bias audit is current and published, then give covered candidates the required notice on time.

Those controls are often collapsed into one “AI transparency” task. New York City’s Local Law 144 does not collapse them. The audit tests disparities in the tool’s outputs, while the notice tells a candidate that an automated employment decision tool will be used and identifies the qualifications or characteristics it will assess. Different teams may produce the evidence, and each obligation has its own deadline.

This explainer describes a compliance workflow based on the law, implementing rules, and city guidance. It is not legal advice.

Start with the decision, not the AI label

Local Law 144 covers an automated employment decision tool, or AEDT, a computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that produces a simplified output used to substantially assist or replace a discretionary hiring or promotion decision.

That definition turns coverage into a workflow question. The Manhattan ranking system produces a score, but the score alone is not enough. An employer must examine how people use it. City rules describe substantial assistance to include relying solely on the output, weighting it more heavily than any other criterion, or using it to overrule conclusions drawn from other factors.

A recruiter who sees a score alongside several equally weighted materials presents a different analysis from a workflow that automatically advances everyone above a threshold. Product descriptions rarely resolve that difference. The useful evidence sits in configuration files, recruiter instructions, approval rules, and logs showing whether a person could advance a low-ranked candidate without an exception.

Some automated systems fall outside this frame because they do not assess candidates for hiring or promotion. A tool that schedules interviews, stores resumes, or transcribes a call may use AI without issuing the kind of decision output covered by the law. A generative model that summarizes resumes is harder to classify from its name alone: if recruiters use its recommendation to screen applicants, the operational role matters more than the vendor’s “copilot” branding.

Geography also needs a recorded decision. New York City’s Department of Consumer and Worker Protection says its rules apply to jobs located in the city at least part time and to certain fully remote roles associated with a city office. The rules define a candidate as someone who has applied for a specific position, which helps separate applicant screening from broad searches for potential prospects.

The notice is a delivery control

For a covered candidate or employee who resides in New York City, the law requires notice at least 10 business days before use of the AEDT. The notice must say that an automated employment decision tool will be used in connection with the assessment or evaluation and state the “job qualifications and characteristics” the tool will assess.

For the Manhattan role, “we use technology in recruiting” is not an adequate operational substitute. The notice should correspond to what the ranking system evaluates, such as experience or skills represented in the application, without making claims that conflict with the tool’s configuration. If the vendor changes the evaluated fields, the notice inventory needs review even when the model keeps the same product name.

City rules allow candidate notice through the employer’s careers website, in a job posting, or by mail or email, subject to the timing requirement. Existing employees may be notified through a written policy or procedure. A defensible record therefore includes the exact text, the publication or delivery channel, the relevant job, and a timestamp that can be compared with the first automated assessment.

The law also gives candidates a route to request an alternative selection process or accommodation. That language does not itself require an employer to offer an alternative process, according to city guidance, although other disability and employment laws may create separate duties. Employers should avoid promising either an automatic alternative or an automatic refusal in a generic notice.

Data disclosure sits beside the main AEDT notice. Information about the type and source of data collected, plus the employer’s data-retention policy, must be available on the website or provided after a written request within the period specified by the law, subject to stated exceptions. That is another reason a notice screenshot alone is incomplete evidence.

The audit is a measurement control

The separate use restriction is explicit: a covered AEDT cannot be used unless “a bias audit has been conducted” no more than one year before use and information about that audit is publicly available. The audit is an impartial evaluation by an independent auditor, not an internal fairness review renamed for procurement.

Under the implementing rules, independence turns on relationships and interests, including whether the auditor helped use, develop, or distribute the tool and whether prohibited employment or financial ties exist. An employer should request an independence statement that addresses those criteria rather than accepting a report whose author is merely described as an outside consultant.

The audit measures how the tool treats specified sex and race or ethnicity categories, including intersectional groups. Depending on whether the system selects candidates or assigns scores, the rules call for selection or scoring rates and impact ratios, which compare a group’s rate with that of the group receiving the highest rate. Categories excluded from calculations need an explanation.

These figures do not prove that a tool is unbiased, job-related, or lawful under every employment rule. An impact ratio can reveal a disparity in the audited data, but it cannot explain whether resume wording reflects unequal access to prior jobs, whether an accommodation failed, or whether a recruiter applies the score differently after the automated stage. Local Law 144 also does not set a passing ratio that turns the audit into a certification.

Data provenance matters. An audit may use historical data from actual operation or test data where permitted, with required disclosures about the source and rationale. If a vendor offers one audit across customers, the employer still needs to establish why that report applies to its deployed system and circumstances. A generic PDF for a product family is weak evidence when the Manhattan workflow uses different thresholds, features, or output logic.

The employer must make a summary of the audit results publicly available before use, along with required information such as the audit date and distribution date. City rules also address how long that material remains posted. Saving the report internally does not satisfy publication, while publishing it does not satisfy candidate notice. These are separate receipts.

Build one file around the Manhattan screening run

A practical review can follow the first applicant through the system and attach evidence at each control point.

1. Map the decision path. Record the job location, who has applied, the system version or configured service, the inputs it reads, the output it produces, and the rule that advances or rejects a person. Ask the recruiting team to demonstrate what happens when a recruiter disagrees with the score.

That establishes whether the output substantially assists the decision instead of assuming every AI-enabled product is covered.

2. Match the audit to the deployment. Obtain the complete audit or available summary, its date, the public URL, the auditor’s independence statement, the data source, the categories analyzed, any exclusions, and the relevant selection or scoring calculations. Compare the tested setup with production.

Where a vendor cannot identify which configuration was audited, pause the claim of coverage rather than treating the product name as proof.

3. Reconstruct notice timing. Preserve the candidate-facing text and evidence of posting or delivery, then compare its timestamp with the moment the system first evaluated each covered candidate. Review whether the named qualifications match the tool’s inputs and whether instructions for requesting an alternative process or accommodation are present.

Ten business days is a scheduling constraint, not wording that can be added after screening.

4. Set two renewal triggers. The audit needs attention before it becomes more than one year old, and material changes to the tool or workflow should trigger a scope review because the existing report may no longer describe production. Notices require their own change control when assessed characteristics, delivery channels, jobs, or geography change.

One calendar reminder cannot manage both obligations.

Return to the Manhattan applicant. The evidence packet should show that the ranking tool was covered, an applicable audit was current and public before use, and the candidate received accurate notice at least 10 business days before the score affected screening. If any link is absent, a vendor assurance form does not reconstruct it.

Questions people ask

Does sending an AI notice make a hiring tool compliant?

No. Notice addresses what a covered candidate or employee must be told and when. The employer must separately establish that an independent bias audit was conducted within the permitted period and that the required summary was publicly available before the tool was used.

Can an employer rely on the vendor’s bias audit?

Potentially, but the report must apply to the tool and deployment, use an auditor who meets the independence requirements, disclose the relevant data and calculations, and remain current. A report covering an unidentified version or materially different configuration does not, by itself, show that the employer’s screening workflow was audited.

Is every resume scanner an automated employment decision tool?

No. Coverage depends on the system’s output and how it affects a hiring or promotion decision. Storage, scheduling, or clerical matching may sit outside the definition, while a score that controls advancement is more likely to qualify. Document the real decision path rather than relying on the product category.

What should an employer request before switching the tool on?

Request the decision-flow documentation, applicable audit materials, public audit link, independence statement, data-source explanation, configuration match, candidate-notice text, and delivery records. For the first Manhattan applicant, those records should establish both clocks before any score changes who receives an interview.