calhire
Market

Hiring in the United States

The United States has had a working test for adverse impact since 1978. What is new is that the tools now being measured against it are automated.

US hiring is governed federally by anti-discrimination law that applies to any selection procedure, automated or not, with adverse impact commonly assessed using the four-fifths rule from the Uniform Guidelines on Employee Selection Procedures. On top of that, New York City requires an annual independent bias audit and candidate notice for automated employment decision tools, and Colorado imposes duties of reasonable care on developers and deployers of high-risk AI systems.

Last reviewed

The short version

  • Federal disparate-impact law does not care whether a screen is a person, a spreadsheet or a model. A selection procedure is a selection procedure.
  • The four-fifths rule is a screening heuristic from the Uniform Guidelines, not a safe harbour and not a legal verdict.
  • NYC Local Law 144 requires an annual independent bias audit, published results and candidate notice for tools in scope.
  • Colorado SB24-205 adds reasonable-care duties and a notification obligation on discovering algorithmic discrimination.

The framework that has been there all along

A great deal of current discussion treats AI hiring regulation as brand new. The federal part is not. Title VII of the Civil Rights Act has prohibited employment practices that produce disparate impact on protected groups without job-related justification for decades, and the Uniform Guidelines on Employee Selection Procedures have given employers a working method for testing it since 1978.

The Uniform Guidelines do not care what produced the selection decision. A written test, a structured interview, an unstructured interview, a physical requirement and a scoring model are all selection procedures, and all are assessed the same way: compare selection rates across groups, and if one group is selected at less than four-fifths the rate of the highest group, that is evidence of adverse impact requiring justification.

This is why an employer adopting automated screening is not entering unregulated territory. They are applying a new instrument to a well-established test, and the practical question is whether they can produce the selection-rate data when asked.

The newer, more specific obligations

NYC Local Law 144
Employers using an automated employment decision tool for candidates in New York City must have it independently bias-audited within the preceding year, publish a summary of the results, and notify candidates that the tool is in use and what it assesses. The audit must be independent, which means the vendor cannot perform it.
Colorado SB24-205
Places duties of reasonable care on developers and deployers of high-risk AI systems, which includes systems making or substantially factoring into consequential employment decisions, together with an obligation to notify the Attorney General on discovering algorithmic discrimination.
Illinois and Maryland
Both regulate specific practices in video interviewing, covering consent, disclosure and the use of facial analysis. CalHire’s AI interview is text-only and no model scores video, which sidesteps this category rather than managing it.
State pay transparency laws
A growing number of states and cities require salary ranges in job postings. This is a posting obligation rather than a screening one, but it is the rule most commonly missed by employers hiring across state lines.

What the platform gives you for this

  • Selection rates, continuously

    Four-fifths ratios computed on disclosed cohorts of at least 30, with groups under 5 reported but never flagged, so a finding is a finding rather than statistical noise.

  • A human in every decision

    No score auto-rejects anyone. Advancing and declining are always attributed to a named person, which is the single most useful fact in a disparate-impact inquiry.

  • Audit-ready data

    Your independent auditor gets the data in the shape they need rather than an export project. We support the audit; we do not perform it.

  • Candidate notice

    Candidates are told an automated tool is in use, what it assesses, and how they can appeal a score. Notice is a requirement in some jurisdictions and good practice in all of them.

Primary sources

What this does not do

Passing the four-fifths rule does not establish that a process is lawful. It is a screening heuristic, and the EEOC has been explicit that meeting it does not guarantee a practice will not be found to have disparate impact. Intersectional effects and small groups below the flagging floor both sit comfortably behind a passing ratio.

We cannot be your independent bias auditor. Local Law 144 requires independence, and a vendor auditing its own tool would defeat the purpose of the requirement. What we can do is make the engagement cheap by having the data ready.

State and city law in this area is moving quickly, and a page like this ages. Treat the linked sources as current and this summary as a starting point. Nothing here is legal advice.

Federal contractors carry additional affirmative-action and record-keeping obligations that this page does not attempt to cover. If you are one, your obligations are broader than what is described here.

Questions people actually ask

Is a skills assessment an automated employment decision tool?
It depends on the jurisdiction’s definition and on how much the output influences the decision. Rather than argue the boundary, the platform is built to satisfy the requirements: candidate notice, human decision-making, and support for an independent bias audit with published results.
What is the four-fifths rule?
Compare each group’s selection rate to the highest group’s. A ratio below 0.8 is evidence of adverse impact requiring justification. It comes from the Uniform Guidelines and is a screening heuristic rather than a legal test in itself.
Who performs the Local Law 144 bias audit?
An independent auditor you engage. The law requires independence, so the vendor cannot do it. We provide the data in the shape the auditor needs and track when the published audit is due for renewal.
Do you use video analysis in interviews?
No. The AI interview is text-only and no model scores video, faces or voices at any stage. This avoids the Illinois and Maryland video-interview requirements rather than managing them.
Can we collect demographic data for measurement?
Yes, on a voluntary basis and held separately from any scoring path. You cannot measure disparate impact without the categories, and letting them reach the decision is the exact harm the measurement is meant to detect.
What if we discover adverse impact?
It surfaces as a signal for investigation. The usual questions are which stage produced the disparity, whether the cohort is large enough to be meaningful, and whether the requirement driving it is genuinely job-related. Colorado additionally imposes a notification obligation on discovering algorithmic discrimination.

Related

Where this connects to the rest of the platform.

  • AI governance, bias auditing and the audit trail

    Adverse-impact analysis on the four-fifths rule, a hash-chained audit trail, candidate appeals, and DSAR handling. The evidence exists before anyone asks for it.

  • Hiring for the public sector

    The sector that invented structured, documented, challengeable selection. Verified assessment is a continuation of that tradition rather than a departure from it.

  • Hiring in the United Arab Emirates

    Emiratisation quotas with real monthly contributions attached, plus a federal data-protection law. National status is a reporting attribute here, never a screening one.

See a verified pipeline for one of your roles

Post a role free and review anonymous, skill-ranked candidates. No card, no sales call to get started.

Start hiring freeTalk to the team