Automated employment decision tool (AEDT) obligations mapped to SWT3 witness procedures. Bias audits, disparate impact analysis, candidate notification, and annual compliance evidence.
Who this is for: HR technology vendors building or deploying automated employment decision tools, employment compliance officers at organizations using AEDTs for hiring or promotion, independent bias auditors conducting annual audits, and legal counsel advising on NYC employment AI compliance.
Actively enforced. NYC Local Law 144 has been in effect since July 5, 2023. Enforcement is by the NYC Department of Consumer and Worker Protection (DCWP). Penalties range from $500 for a first violation to $1,500 per violation per day for subsequent violations. Each candidate screened without a compliant bias audit or proper notice may constitute a separate violation.
NYC Local Law 144 (Int. 1894-2020) is the first US municipal law to regulate the use of artificial intelligence in employment decisions with active enforcement. The law requires employers and employment agencies that use automated employment decision tools (AEDTs) to meet three core obligations:
The law is notable for several reasons. It was the first US municipal AI regulation with active enforcement. It applies to both employers and employment agencies. And its definition of AEDT -- any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that substantially assists or replaces discretionary decision-making -- has influenced subsequent state and federal AI proposals.
Under the DCWP's final rules, an AEDT is any computational process that meets two criteria: (1) it is derived from machine learning, statistical modeling, data analytics, or artificial intelligence, and (2) it substantially assists or replaces discretionary decision-making in employment decisions.
"Substantially assists" means the tool produces a score, classification, recommendation, or ranking that is used to make or meaningfully influence an employment decision. The following tool categories qualify:
Tools that do not substantially assist or replace discretionary decision-making are excluded. Examples include:
The distinction centers on whether the tool's output substantially assists or replaces the human decision. If a recruiter routinely follows the tool's recommendations without independent evaluation, the tool qualifies as an AEDT even if a human technically makes the final decision.
The bias audit is the central compliance mechanism of LL 144. The DCWP's final rules define specific requirements:
The auditor must be independent of the employer and the AEDT vendor. The DCWP does not prescribe specific auditor qualifications, but the auditor cannot be the entity that developed or deployed the tool. An auditor employed by the same organization that developed the AEDT is not independent.
The audit must calculate disparate impact using selection rates (for binary pass/fail tools) or scoring rates (for tools that produce scores or rankings). The analysis must cover:
The standard metric is the impact ratio: the selection rate for each category divided by the selection rate for the most-selected category. A ratio below 0.8 (the four-fifths rule from EEOC Uniform Guidelines) is a common benchmark for indicating adverse impact, though LL 144 does not mandate a specific threshold.
The audit must use historical data from the AEDT's actual use. If insufficient historical data is available (e.g., a new tool), the audit may use test data, but this must be disclosed in the published summary. The number of individuals assessed must be reported for each category.
The summary must be posted on the employer's or employment agency's website and must remain available for at least six months after the AEDT's last use. The summary must include the date of the audit, the source and explanation of data used, and the selection or scoring rates for each demographic category.
Each LL 144 obligation maps to SWT3 witness procedures that produce cryptographically anchored evidence of compliance activity.
| LL 144 Requirement | SWT3 Procedure | What It Witnesses | Evidence Produced |
|---|---|---|---|
| Disparate impact monitoring | AI-FAIR.1 | Demographic parity measurement across protected categories | Factor A: metric name, Factor B: measured value, Factor C: threshold comparison |
| Protected class identification | AI-FAIR.2 | Detection and documentation of protected class categories in the analysis | Factor A: detection method, Factor B: categories identified, Factor C: data source |
| Bias mitigation steps | AI-FAIR.3 | Mitigation actions taken when disparate impact is identified | Factor A: mitigation method, Factor B: pre-mitigation metric, Factor C: post-mitigation metric |
| Human review of automated decisions | AI-HITL.1 | Human override capability and exercise for employment decisions | Factor A: override authority, Factor B: review scope, Factor C: override rate |
| Annual audit evidence chain | AI-AUDIT.1 | Audit trail integrity for the independent bias audit | Factor A: log source, Factor B: integrity hash, Factor C: retention period |
| Candidate notification content | AI-EXPL.1 | Explanation of what the tool assesses and how candidates can request alternatives | Factor A: explanation method, Factor B: confidence, Factor C: factors cited |
| Automation scope tracking | AI-AUTO.1 | Boundary between automated and human decision-making in the employment process | Factor A: automation boundary, Factor B: decision type, Factor C: human involvement level |
| Training data governance | AI-DATA.1 | Data provenance and governance for the datasets used in the bias audit | Factor A: data source, Factor B: data hash, Factor C: governance status |
LL 144 requires: The bias audit must calculate selection rates or scoring rates for each race/ethnicity and sex category, including intersectional combinations. The impact ratio (category rate divided by most-selected category rate) must be computed and published.
How SWT3 addresses it: The witness_fairness() call captures the specific metric being measured (e.g., selection rate for Hispanic or Latino Female), the measured value (e.g., 0.72), and the threshold comparison (e.g., impact ratio 0.85 against most-selected category). Each demographic category produces a separate anchor, creating a complete record of the disparate impact analysis.
Filter the witness ledger for AI-FAIR.1 anchors grouped by audit date. Each anchor represents one demographic category's selection or scoring rate. Factor A identifies the metric and category. Factor B contains the measured rate. Factor C shows the impact ratio against the most-selected group. Verify that all required race/ethnicity, sex, and intersectional categories are present. Gaps in category coverage indicate an incomplete audit.
LL 144 requires: While LL 144 does not mandate specific remediation when disparate impact is found, employers must be prepared to demonstrate what actions were taken if the DCWP investigates. Documenting mitigation steps is a best practice that reduces enforcement risk.
How SWT3 addresses it: The witness_bias_mitigation() call records the mitigation method applied (e.g., threshold adjustment, feature removal, retraining), the pre-mitigation metric, and the post-mitigation metric. This creates a before-and-after evidence chain showing that the employer responded to identified disparate impact.
AI-FAIR.3 anchors should follow AI-FAIR.1 anchors chronologically, demonstrating that mitigation was performed after disparate impact was identified. Factor B (pre-mitigation metric) should match the corresponding AI-FAIR.1 Factor B value. Factor C (post-mitigation metric) should show improvement. If no AI-FAIR.3 anchors exist after AI-FAIR.1 anchors showing adverse impact ratios, this is a finding.
LL 144 requires: Candidates must be informed of how to request an alternative selection process or a reasonable accommodation. This implies that a human review pathway must exist as an alternative to the automated process.
How SWT3 addresses it: The witness_human_override() call records the override authority (who can intervene), the scope of review (which decisions are subject to human review), and the override rate (how frequently the human path is used). This documents that the alternative process exists and is operational.
AI-HITL.1 anchors prove the alternative selection pathway exists. Factor A identifies who has override authority. Factor C (override rate) indicates whether the alternative process is genuinely available or merely nominal. A zero override rate combined with a high volume of automated decisions may indicate the alternative process is not meaningfully accessible.
LL 144 requires: Notice to candidates must include the job qualifications and characteristics that the AEDT will assess. Candidates must understand what the tool evaluates and how it influences the employment decision.
How SWT3 addresses it: The witness_explanation() call records the explanation method (e.g., published notice, email disclosure), the content provided, and the specific factors cited. This creates a timestamped record of what candidates were told about the tool's assessment criteria.
AI-EXPL.1 anchors must predate AI-INF.1 anchors (actual AEDT use) by at least 10 business days, proving that notice was provided before the tool was applied to the candidate. Factor C (factors cited) should match the qualifications and characteristics the AEDT actually assesses. A mismatch between the disclosed factors and the tool's actual assessment criteria is a compliance gap.
LL 144 requires: The law applies only to tools that "substantially assist or replace" discretionary decision-making. Employers must understand and document the boundary between automated and human decision-making in their hiring process.
How SWT3 addresses it: The witness_automated_decision() call records the automation boundary (what the tool decides versus what humans decide), the decision type (screening, ranking, selection), and the level of human involvement. This creates evidence of the employer's assessment of whether the tool qualifies as an AEDT.
AI-AUTO.1 anchors document the employer's determination of AEDT scope. Factor A defines the automation boundary. Factor C describes human involvement. If Factor C indicates minimal human involvement (e.g., rubber-stamping automated rankings), the tool likely qualifies as an AEDT even if the employer classifies it otherwise. Cross-reference with AI-HITL.1 override rates to validate the stated human involvement level.
| Examiner Question | Where to Look |
|---|---|
| Has an independent bias audit been conducted within the past year? | AI-AUDIT.1 anchors. The most recent anchor timestamp must be within 365 days of the AEDT's current use. Factor B (integrity hash) verifies the audit record has not been altered. |
| Does the audit cover all required demographic categories? | AI-FAIR.1 anchors grouped by audit date. Count unique categories in Factor A. Must include all 7 race/ethnicity categories, 2 sex categories, and intersectional combinations. Missing categories indicate an incomplete audit. |
| What are the selection rates and impact ratios? | AI-FAIR.1 Factor B (measured rate) and Factor C (impact ratio). Ratios below 0.8 against the most-selected category indicate potential adverse impact under the four-fifths rule. |
| Were candidates notified at least 10 business days in advance? | Delta between AI-EXPL.1 timestamp (notice provided) and AI-INF.1 timestamp (tool applied). Must be at least 10 business days. Factor C in AI-EXPL.1 lists the qualifications disclosed. |
| Is there an alternative selection process available? | AI-HITL.1 anchors prove the alternative pathway exists. Factor C (override rate) indicates whether it is genuinely accessible. Zero override rate with high automated volume is a red flag. |
| Are audit results publicly posted? | AI-EXPL.1 or AI-AUDIT.1 anchors documenting the publication. Factor A should indicate the posting location (employer website URL). Results must remain posted for 6 months after the AEDT's last use. |
| What data was used in the bias audit? | AI-DATA.1 anchors. Factor A identifies the data source (historical use data vs. test data). Factor C shows governance status. If test data was used, the published summary must disclose this. |
| What mitigation was performed after adverse impact was identified? | AI-FAIR.3 anchors following AI-FAIR.1 anchors that show impact ratios below 0.8. Factor B (pre-mitigation) and Factor C (post-mitigation) document the remediation effect. |
Full SDK documentation: sovereign.tenova.io/docs
Create a free account: sovereign.tenova.io/signup