By: Dave Schmidt and Bre Timko
On April 6, 2023, the New York City (NYC) Department of Consumer and Worker Protection (DCWP) issued the final set of rules for enforcing Local Law 144, regulating the use of automated employment decision tools (AEDTs) wherein hiring and promotion decisions are made or substantially assisted by algorithmically-driven mechanisms. The enforcement schedule has been pushed back to July 5, 2023. Taken together, the law and the rules finalize the steps employers and employment agencies need to take when using an AEDT in NYC after the enforcement date.
Given the broad-spanning nature of this law, the second set of proposed rules provided by the DCWP drew almost as much commentary (175 pages of written commentary and 40 minutes of verbal commentary at the public hearing) as the first set of proposed rules. The vast majority of the proposed rules stayed the same or were changed with no significant impact on the steps required to come into compliance. However, a handful of changes do, in fact, affect bias audit methodology and publication requirements.§5-300. Definitions:
- 14 of the 17 definitions in the final rules were unchanged; 2 definitions had very minor changes.
- The only significant change was to the definition of “Machine learning, statistical modeling, data analytics, or artificial intelligence” where the 3rd prong (“for which the inputs and parameters are refined through cross-validation or by using training and testing data”) was removed; this serves to eliminate some specific situations in which the law may have been interpreted as not being applicable.
- No change to the AEDT definition, nor the “substantially assists” clause within that definition.
- No clarification on what constitutes “test data” – it is still broadly defined as “data used to conduct a bias audit that is not historical data.”
- Added §5-301(b)(5) and §5-301(c)(5) requiring that the audit state the number of individuals assessed by the AEDT who did not provide demographic data.
- Added §5-301(d) allowing an auditor to exclude from analysis a demographic group that is less than 2% of the sample, however, the auditor must, 1) provide justification for the exclusion, and 2) still report the number of applicants and the scoring rate or selection rate for any excluded group.
- Added context to the selection rate example scenario that makes it clear that a bias audit must be conducted if an AEDT is used in any part of the process (e.g., to screen), not just if it is part of the final hiring decision.
- Updated scoring rate tables, which now include the number of applicants – in some places, the sample sizes are very small (e.g., N=4); any selection rates and impact ratios calculated based on very small samples should be viewed cautiously.
- This section added 3 use cases to help clarify the data that may be used in an audit.
|
Use Case (Examples) |
Bias audit using specific employer historical data |
Bias audit using historical data from other employers |
Bias audit using “test data” |
|
1) Prior to use of the AEDT |
Not possible |
Yes |
Yes |
|
2) AEDT in use, but insufficient employer historical data |
Not possible |
Yes, but only if employer historical data are included in analyses |
Yes |
|
3) AEDT in use and sufficient employer historical data |
Yes |
Yes, but only if employer historical data are included in analyses |
No |
§5-303. Published Results:
- This section added language to §5-303(a) that published results tables must now include the “number of applicants or candidates … for all categories.”
- Consistent with the language added to §5-301(b)(5) and §5-301(c)(5), the published results tables must also include a note that indicates “the number of individuals the AEDT assessed that fall within an unknown category” (i.e., missing demographic data).
- Changed an “or” to an “and” in the three prongs in §5-304(d) that describe how to comply with §20-871(b)(3), increasing the compliance requirements for employers and employment agencies.
We recommend that employers evaluate if any assessments used in NYC would be covered by this law. If there are such AEDTs and the bias audit has not yet been started, these projects should start immediately if the goal is to meet the July 5 enforcement date. Bias audits require time to conduct properly and are more complex than many realize. If the bias audit has already been conducted anticipating the April 15 enforcement, these final rules should be reviewed with the auditor to identify what changes need to be made to the data, the analyses, and/or the reports previously produced.
For better or worse, the rules for NYC Local Law 144 are now final, and a further stay of enactment is not anticipated. Please feel free to reach out to DCI to discuss any NYC Local Law 144 bias audit needs in more detail.
