"Automated employment decision tool" started as a legal term in New York City and has become the general name for any software that screens, scores or ranks candidates. If your team uses one, three questions follow: does it count under the rules where you hire, could it be biased, and how would you know?
This is a plain-English explainer for talent teams, not legal advice. Rules differ by jurisdiction and are still changing, so confirm your position with counsel. For the full obligations law by law, see our guide to AI hiring laws in 2026.
What is an automated employment decision tool?
An automated employment decision tool (AEDT) is a computer-based tool that uses machine learning, statistical modelling, data analytics or artificial intelligence to produce a simplified output, such as a score, classification or recommendation, that substantially assists or replaces discretionary decisions about hiring or promotion. That is the definition in New York City's Local Law 144 of 2021, the law that made the acronym common.
New York City's rules add detail that is useful anywhere:
- Simplified output includes a score (rating a candidate's technical skills), a tag or category (sorting CVs by keywords), a recommendation (whether to interview someone) or a ranking (ordering candidates by how well their cover letters match the job description). Tools that only translate or transcribe, such as converting a CV from PDF or transcribing an interview, do not produce simplified output.
- Substantially assist or replace means any one of three things: you rely solely on the output; you use it as one of several criteria and it carries more weight than any other; or you use it to overrule conclusions drawn from other factors, including human judgement.
- Exclusions in the law itself include junk email filters, firewalls, antivirus software, calculators, spreadsheets, databases and data sets.
How do US and EU rules define automated hiring tools?
Each regime uses a different name and a different trigger, but all of them reach the tools that score or rank applicants. The broadest definitions also cover job-ad targeting and tools that merely help a person decide.
| Jurisdiction | Term used | What is covered | Main duties | Applies from |
|---|---|---|---|---|
| New York City (Local Law 144) | Automated employment decision tool | Tools that substantially assist or replace hiring or promotion decisions about candidates for a specific position | Independent bias audit in the past year, published summary, notice 10 business days before use | Enforced since 5 July 2023 |
| California (Civil Rights Council regulations) | Automated-decision system | A computational process that makes a decision or facilitates human decision-making about an employment benefit | Using one that discriminates is unlawful; anti-bias testing counts as evidence; keep its data four years | 1 October 2025 |
| Colorado (SB 26-189) | Automated decision-making technology | Technology that processes personal data and uses computation to generate output, such as predictions, scores or classifications, used to make, guide or assist a consequential decision, employment included | Notice, explanation within 30 days of an adverse outcome, data correction, human review, records for three years | 1 January 2027 |
| European Union (AI Act) | High-risk AI system (Annex III, point 4) | AI used to target job ads, analyse and filter applications, evaluate candidates, and for promotion, termination, task allocation and monitoring | Human oversight, logs kept at least six months, inform workers and candidates | 2 December 2027 |
| US federal law | No AI-specific term | Any selection procedure, automated or not | Title VII disparate impact; the four-fifths rule of thumb in the Uniform Guidelines | Already in force |
Three differences matter in practice.
California's definition is the widest. It names computer-based tests, puzzles and games, targeted job advertising, CV screening for terms or patterns, analysis of facial expression, word choice or voice in online interviews, and analysis of data bought from third parties. It applies to employers with five or more employees, and it treats a vendor acting on an employer's behalf as an employer too.
New York City's law is narrower on stage. Its FAQ says the requirements apply to assessing people who have applied for a specific position, so scanning a CV database, outreach and inviting applications are outside it. It covers any job based at a New York City office, including fully remote jobs associated with one.
The EU AI Act reaches beyond the EU. It applies to deployers in the EU and to providers and deployers elsewhere when the system's output is used in the EU. Separately, AI that infers emotions at work has been banned since 2 February 2025, which matters for any video interview product that claims to read emotion.
Reviewing the tools in your hiring process?
Cohesyve takes a job description, finds candidates, reaches out to them and runs an AI job simulation of the real work, then returns a ranked shortlist. If you are looking at tools, it is one to consider.
Which hiring tools usually count as an AEDT?
Any tool whose output decides or heavily shapes who moves forward is likely in scope. Tools that move information around without judging it usually are not.
| Tool type | Likely in scope? | Why |
|---|---|---|
| CV screening or ranking, including keyword filters | Yes | Produces a tag, score or ranking that decides who is reviewed |
| Auto-scored skills tests, coding tests and job simulations | Yes, if the score drives who advances | A score that carries the most weight meets New York City's second prong |
| Game-based and psychometric assessments | Yes | Named in California's definition; the EU covers tools that evaluate candidates |
| One-way video interviews with AI analysis | Yes | California names face, voice and word-choice analysis; EU bans emotion inference at work |
| Chatbots that apply knockout questions | Often | Rejecting on an automated rule is relying solely on the output |
| Sourcing and job-ad targeting | Depends on the law | Outside New York City's law; inside California's and the EU's |
| Scheduling, transcription and ATS storage | Usually no | They do not judge candidates; New York City excludes transcription |
A general-purpose AI model can become an AEDT through how you use it. If a recruiter pastes CVs into a chatbot and shortlists whoever it ranks highest, the definitions look at that use, whatever the product was built for.
Does a human reviewer take a tool out of scope?
Usually not. New York City counts a tool as in scope when its output outweighs every other criterion or overrules human judgement, and California covers systems that "facilitate" human decisions as well as those that make them.
The EU AI Act does let a provider argue that an Annex III system is not high-risk if it only performs a narrow procedural or preparatory task and does not materially influence the outcome. That exemption never applies to a system that profiles people, and scoring candidates on their personal characteristics will often be profiling. If a hiring manager only ever sees the candidates the tool put at the top, the tool is making the decision.
Where does AI hiring bias come from?
AI hiring bias comes from what a tool learns from and what it measures. A model trained on past decisions copies the patterns in them, and a model that measures things unrelated to the job can screen out groups without anyone intending it.
The common sources:
- Historical training data. Amazon built a CV-ranking tool trained on the CVs of its existing, mostly male, engineering workforce. As reported by Reuters in 2018, it learned to penalise CVs containing the word "women's" and graduates of certain women's colleges, and Amazon stopped using it.
- Proxies. California's regulations define a proxy as a characteristic closely correlated with a protected basis. Postcodes, employment gaps, graduation years and the names of schools and clubs can all stand in for race, sex, age or disability.
- Names and language in large language models. A University of Washington study presented in October 2024 ran more than three million comparisons of CVs against job listings with three language models. White-associated names were preferred 85% of the time and Black-associated names 9%; the models never preferred Black male names over white male names.
- Measuring the wrong thing. A tool that scores tone of voice, typing speed or game reaction times may measure traits the job does not need. If a measure is not job-related, any group difference it creates is hard to defend. Our guides to criterion-related validity and content validity explain how to check.
- Accessibility. A timed, camera-on or audio-based format can screen out disabled candidates who could do the job. You need an accommodation route regardless of how the tool performs overall.
- Feedback loops. A tool retrained on whom you hired never learns about the people it rejected, so an early skew can grow over time.
Human screening has its own biases, which our article on unconscious bias in recruitment covers. The difference with a tool is scale: one skewed model applies the same skew to every applicant.
How do you measure bias in an automated hiring tool?
Calculate an impact ratio for each group: that group's selection rate divided by the selection rate of the most selected group. The US Uniform Guidelines treat a ratio below four-fifths (80%) as general evidence of adverse impact, and add that smaller differences can still count when they are statistically and practically significant.
A hypothetical example: if 50% of men and 35% of women who take an assessment pass it, the impact ratio for women is 0.35 ÷ 0.50 = 0.70, below the four-fifths line. For tools that output a score instead of pass or fail, New York City's rules use a scoring rate, the share of each group scoring above the sample's median score, in the same way.
New York City's audit rules set useful norms even outside the city:
- The audit uses data from your actual use of the tool. Where there is not enough, data from other employers using the same tool or test data can be used, with an explanation.
- Demographics cannot be imputed or inferred by software. They come from what candidates report.
- Categories under 2% of the data can be left out of the calculations, with a justification.
- The auditor must be independent of both the employer and the vendor.
One gap is easy to miss. New York City's law requires the audit but does not require any action based on its results. Federal, state and city anti-discrimination law still applies, so a poor impact ratio that you published and ignored is a problem.
How should employers evaluate an AEDT before buying?
Ask for evidence on three things: what the tool measures, how it performs across groups, and what the vendor will hand you for your own compliance. The table below sets out what a good answer looks like.
| Ask the vendor | A good answer | A red flag |
|---|---|---|
| What exactly does the tool measure, and why is it job-related? | Named skills tied to the job description, with a validity study | "Our model finds top performers" with no detail |
| When was the last independent bias audit, and on whose data? | Dated within the past year, auditor named, impact ratios by sex, race/ethnicity and intersections | An internal review, or an audit more than a year old |
| Can a reviewer see why a candidate scored as they did? | Scores broken down against a visible rubric or the candidate's own work | One number with no breakdown |
| Does it use video, voice or facial analysis? | No emotion inference; any analysis disclosed and optional | Claims to read personality or emotion from video |
| What accommodations can candidates request? | Extra time, alternative formats, a human-reviewed route | None, or "contact support" |
| What documentation will you provide for the EU, California and Colorado? | Intended uses, known limitations, human-review guidance, data retention | "We are compliant" without documents |
Our guide to choosing a pre-employment assessment platform has a fuller buyer's checklist. Formats that score a sample of the real work are easier to defend because the link to the job is visible; see what a work sample test is. If you are weighing video tools, read our explainer on one-way video interviewing first.
How do you govern automated employment decision tools once they are live?
Treat each tool like a selection procedure you own: list it, give it an owner, tell candidates, keep a person able to overrule it, and check its results on your own applicants on a schedule. The vendor's audit does not cover how the tool behaves on your roles and your applicant pool.
- Inventory. List every tool that touches a hiring or promotion decision, what it outputs and how much weight that output carries.
- Map where candidates are. Note any roles tied to New York City, California, Colorado or the EU, including remote ones.
- Notify. New York City needs notice 10 business days before use, with how to request an accommodation. Colorado needs clear notice at the point of interaction from 1 January 2027. From 2 December 2027 the EU requires workers' representatives and affected workers to be told before a high-risk system is used at work, and candidates to be told when one is used in decisions about them.
- Keep a human able to intervene. Name reviewers, train them on what the tool measures, and give them authority to override it. Record when they do.
- Monitor impact ratios. Recalculate on your own data at least yearly, and after any change to the tool or the role.
- Offer an alternative. A route for candidates who need an accommodation or who ask for human review of a rejection.
- Keep records. Four years for California, including the tool's data; three years for Colorado; at least six months of system logs under the EU AI Act; a fresh audit every year for New York City.
For a structure to hang this on, the US National Institute of Standards and Technology publishes a voluntary AI Risk Management Framework built on four functions: Govern, Map, Measure and Manage. Pair any tool with a structured interview, so the human stage of your process is as consistent as the automated one.
Where does Cohesyve fit?
Cohesyve is one of these tools, so we apply this guide to ourselves. A company gives Cohesyve a job description; Cohesyve sources candidates, does outreach, puts them through AI job simulations of the real work and returns a ranked shortlist. A ranked shortlist is the kind of output these rules cover, so we expect customers to treat it as in scope: candidates see the work they are assessed on, and the hiring team reviews the shortlist and decides who moves forward. Plans are a subscription from $199/month billed annually, and new accounts get 10 free candidates.
Common questions
What is an AEDT in simple terms? Software that gives a candidate a score, rank, tag or recommendation that heavily shapes whether they move forward. The acronym comes from New York City's Local Law 144.
Who is responsible for the bias audit, the vendor or the employer? Under New York City's law, the employer or employment agency. A vendor can commission an independent audit of its tool, but the employer must make sure one exists before using the tool.
How often is a bias audit needed? Every year in New York City. An audit can be relied on for one year from the date it was carried out.
What are the penalties under Local Law 144? Up to $500 for a first violation and any others on the same day, then $500 to $1,500 for each subsequent one. The city's consumer protection department enforces the audit and notice rules; discrimination claims go to the city's Commission on Human Rights.
Does a skills test count as an automated employment decision tool? If it is scored automatically and the score drives who advances, usually yes. California names computer-based tests explicitly.
Does the EU AI Act apply to a US employer? It can. The Act covers deployers outside the EU when the AI system's output is used in the EU, for example when hiring for an EU-based role.
Is AI hiring bias worse than human bias? A biased tool repeats the same skew for every applicant, so it reaches more people than one biased interviewer. A tool's results can also be measured group by group, which is hard to do with individual interviewers.