NYC Local Law 144: What Counts as an AEDT

Nearly every dispute about Local Law 144 is a dispute about whether the thing you bought is an AEDT. The statute gives four elements and the 2023 rules give a three-branch test for the hardest of them.

The statutory definition, in parts

Administrative Code section 20-870 defines an automated employment decision tool as any computational process, derived from machine learning, statistical modelling, data analytics or artificial intelligence, that issues simplified output — including a score, classification or recommendation — that is used to substantially assist or replace discretionary decision making for making employment decisions that impact natural persons.

Four elements have to be satisfied together:

• A computational process derived from machine learning, statistical modelling, data analytics or AI. The DCWP rules gloss this as techniques for which a computer identified the inputs, the relative importance placed on those inputs, and any other parameters, in order to improve the accuracy of a prediction or classification.
• It issues a simplified output — a prediction or classification, which may take the form of a score, tag, categorisation, ranking or recommendation.
• The output substantially assists or replaces discretionary decision making.
• For an employment decision, meaning screening candidates for employment or employees for promotion within New York City.

The first element is narrower than “software”. A rules engine a human wrote — reject anyone without a required licence — is not a computational process for which a computer identified the weights. It is a filter, and the rules’ gloss on parameter learning is what excludes it. The Department of Consumer and Worker Protection publishes the statute, the final rules and its own guidance at nyc.gov.

The three-branch substantial-assistance test

The April 2023 DCWP rules define “substantially assist or replace discretionary decision making” exhaustively, as any one of three things:

• Relying solely on a simplified output — a score, tag, classification, ranking — with no other factors considered.
• Using a simplified output as one criterion among several, where it is weighted more than any other criterion in the set.
• Using a simplified output to overrule conclusions derived from other factors, including human decision making.

This is where the analysis actually lives, and it is a question about your process rather than about the vendor’s product. The same resume-ranking model is inside branch two in a team that sorts by score and reviews the top twenty, and outside all three branches in a team where a recruiter reads every application and the score is one of five inputs none of which dominates. Nothing about the software changed.

Branch three deserves attention because it is easy to build by accident. A hard cut-off — a field like knockout_score that removes a candidate from the queue before a human sees them — overrules human judgement by definition, since there was no human judgement left to exercise. Teams often describe such a gate as “just a pre-filter”, which is precisely the thing the branch is written to catch.

Simplified output, and what is excluded

The rules exclude from “simplified output” the results of tools that translate or transcribe existing text: converting a resume from PDF, transcribing a recorded video or audio interview, and similar format conversions. The reasoning is that these produce a representation rather than a prediction, and no ranking is implied by them.

A large language model complicates this exclusion in a way the 2023 rules did not anticipate. If a model transcribes an interview, it is excluded. If the same model summarises the interview into three bullet points, it has produced neither a score nor a classification in the ordinary sense — but a summary that consistently foregrounds some candidate attributes and omits others is doing selection work in narrative form. If it is asked to output strong fit / possible fit / weak fit, it has plainly produced a classification and the exclusion does not apply. The middle case is not addressed by the text of the rules.

Applying it to real tools

Working through the categories concretely:

• A resume screener that ranks applicants and whose ranking determines who is reviewed — an AEDT, on branch one or branch two. This is the paradigm case.
• Interview scheduling software that finds mutual availability — not an AEDT. It issues no prediction or classification about a candidate and makes no employment decision.
• A keyword search over a resume database — generally not an AEDT, because the recruiter chose the keywords and no model learned the weights. A semantic search that ranks by learned embedding similarity is a much closer question.
• A video interview tool scoring communication skills — an AEDT if the score is used in screening, and note that the same product triggers separate duties under the Illinois AI Video Interview Act and under Maryland’s facial recognition consent law.
• A promotion model ranking internal candidates — an AEDT, because employment decisions expressly include promotion.
• A chatbot that answers candidate questions — not an AEDT, unless it also scores or routes the candidate, at which point it is.

The geography element is also a real limit and is often overstated in both directions. The law reaches employment decisions for positions within New York City; the widely used practical reading is that a fully remote role with no NYC nexus is outside it, while a hybrid role based in the city is inside it regardless of where the employer or the screening vendor sits.

What is still unresolved

Three questions have no authoritative answer as at the date on this page, and it is more useful to name them than to guess.

Whether the vendor can ever be the regulated party. The duties in sections 20-871 and 20-872 fall on employers and employment agencies. A pure software vendor is not obviously either, which is why the audit-sharing mechanism exists — but an agency-like vendor that itself screens on the employer’s behalf may be an employment agency in substance. That is a fact question nobody has litigated publicly.

Where a narrative LLM assessment falls. As above: the transcription exclusion is clear at both ends and silent in the middle, and the rules predate the general use of generative models in screening.

How much enforcement there will be. Local Law 144 is complaint-driven, and DCWP does not proactively crawl for missing audit summaries. Independent research published after the law took effect reported that only a small minority of covered employers had posted a summary at all. That is an observation about compliance rates, not a licence — the law binds whether or not anyone complains, and the per-day violation structure means a lapse compounds. For what the audit itself must contain, see the bias audit page.