Do NYC employers have to disclose AI CV screening?

What did Mamdani say about AI, and why does it matter to job seekers?

On Monday night, Mayor Zohran Mamdani told New Yorkers that the federal government has abdicated its responsibility on artificial intelligence, and urged the public to push back against the idea that an AI takeover is inevitable — demanding regulation that protects people and returns power to their hands. That's a political statement, not a new rule. But the category of software he's pointing at already has a local rulebook in hiring, and you're probably running into it without being told: automated tools that rank, score or filter job applicants before a human ever opens the file.

New York City has regulated those tools since July 2023, when enforcement of Local Law 144 began. It's the first law in the United States to put a disclosure-and-audit obligation on what it calls automated employment decision tools — AEDTs. For a job seeker, the law's existence changes something concrete: in some NYC hiring processes, you're entitled to know in advance that a machine is scoring you, and to see a published summary of how that machine performed across race and sex categories. Few candidates ever look.

Here's the gap worth understanding before you go hunting for a notice. The law's definition is narrower than most people assume, enforcement has been modest, and a huge share of automated filtering — plain keyword searches inside an applicant tracking system, recruiter-side ranking that a human then reviews — sits outside it. So the practical answer isn't "demand transparency and wait." It's to write your CV on the assumption that software reads it first, every time, whether or not anyone tells you so. That assumption costs you nothing and protects you everywhere.

What exactly does Local Law 144 require employers to tell you?

Local Law 144 requires three things from an employer or employment agency using an AEDT on NYC candidates: a notice at least 10 business days before the tool is used, disclosure of the job qualifications and characteristics the tool assesses, and a publicly available summary of an independent bias audit conducted within the past year. The notice can appear on the careers page, in the job posting itself, or by direct email or mail. Employers must also tell you, on request, what data the tool collects and the source and retention policy for that data — though the law gives them 30 days to answer.

The bias audit is the part most candidates never read, and it's the most revealing. An independent auditor calculates selection or scoring rates by sex, by race and ethnicity, and by intersectional category, then reports impact ratios comparing each group to the most-selected group. The summary has to state the distribution date of the tool and the number of individuals the tool assessed, including any excluded from the calculation because their category fell below a certain size. If an employer posts an audit with huge "unknown" category counts, that tells you something about the quality of their data.

What you cannot get is the one thing you'd actually want: your own score, the weights the model used, or an explanation of why you were filtered out. Local Law 144 is a transparency-and-audit law, not an individual-rights law. There's no right to an explanation, no right to human review, and no right to opt out of the tool and still be considered. The same limitation shows up in other jurisdictions — the UK's position on automated CV screening in 2026 follows a similar pattern, where disclosure exists but opting out isn't realistic.

Which hiring tools actually count as an automated employment decision tool?

An AEDT under the law is a computational process derived from machine learning, statistical modelling, data analytics or AI that issues a simplified output — a score, classification or recommendation — used to substantially assist or replace discretionary decision-making about hiring or promotion. The two load-bearing phrases are "simplified output" and "substantially assist." A tool substantially assists when it's the only criterion, or when its output is weighted more than any other factor, or when it overrules a human conclusion. That's a high bar, and employers read it generously in their own favour.

Run that definition against what you actually encounter. A video interview platform that scores your answers and ranks you: squarely covered. A resume-screening model that assigns each applicant a fit score the recruiter sorts by: covered, if that score drives the cut. A Boolean keyword search inside an applicant tracking system, where a recruiter then reads 60 CVs by hand: almost certainly not covered, because there's no model-derived score doing the deciding. A spell-check on your cover letter: obviously not. Most of the filtering that kills applications in practice lives in that third bucket, unregulated and invisible.

This is why the regulation debate, while worth having, is a poor substitute for preparation. Parsing failures aren't discriminatory decisions requiring an audit — they're mechanical. If your two-column template collapses your job titles into a single garbled line, no bias audit will catch it and no notice will warn you. The bias-audit question across hiring AI is about whether the model treats groups fairly; your immediate problem is whether the model can read you at all.

How do you find the AI disclosure and bias audit for a job you're applying to?

Start on the employer's careers page and look for a link labelled AEDT notice, automated employment decision tool, or bias audit results — it's usually in the footer or a compliance section near the EEO statement. If it isn't there, check the job posting text itself and the application portal's terms page. Many large NYC employers host a single page covering every tool they use, with a PDF audit summary per vendor. Search the employer name alongside "bias audit summary" if the careers page buries it.

Read the audit for three things. First, the assessment count: a summary covering 400 applicants tells you far less than one covering 40,000. Second, the impact ratios — anything materially below 0.80 for a group is the figure regulators and plaintiffs' lawyers watch. Third, the proportion of candidates with unknown demographic category, because audits built mostly on test data or inferred categories are weak evidence of anything. If the audit is older than a year, the employer is out of compliance, and that's a fair thing to raise politely with the recruiter.

Then decide what to do with it. Honestly? For most candidates the answer is: note it, and apply anyway. Filing a complaint with the Department of Consumer and Worker Protection is your right, and employers face civil penalties per violation with each day of a continuing violation treated separately — but it won't get you the interview. The disclosure's real value to you is informational: it tells you a model scores you early, which tells you exactly how much your file's machine-readability matters.

What should you change on your CV when you know an algorithm reads it first?

Three changes do most of the work: mirror the posting's own vocabulary, strip anything that breaks text extraction, and put your evidence in plain numbers. Mirroring means using the employer's exact term — if the ad says "revenue operations," don't write "RevOps" and hope the model generalises. List both forms once each where it reads naturally. Models and keyword filters both reward literal overlap, and the posting is the only reliable source for what that vocabulary is. Copy the responsibilities section into a scratch document and mark which terms genuinely appear in your history.

Formatting is where good candidates lose. Two-column layouts can interleave text from both columns when extracted, so a parser reads your skills list spliced into your job dates. Icons replace words — a phone glyph where "Phone" should be. Text boxes, headers, footers and tables are frequently dropped entirely, which is how people lose their contact details without knowing. Dates need a consistent, unambiguous format: "Mar 2022 – Aug 2024" parses; "Spring '22 to late '24" does not. One column, real section headings, a standard font, a PDF exported from text rather than a scan. If you've built your CV in a design tool, see what the scanner actually reads before you send it.

Evidence is the third fix, and it's the one that survives both machine and human review. Replace "responsible for improving onboarding" with "cut new-hire onboarding from 14 days to 5 across 120 hires." Numbers, scope, timeframe. Then verify the file rather than trusting it — run it through a free CV check that scores ATS compatibility and see your overall score out of 100 plus category scores for experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, with a visual layout analysis showing how the page reads. Paste the job description in and you'll get the missing skills listed, plus ATS-friendly rewrites in six templates that mirror the posting's keywords.

Frequently asked questions