CoChat
Sign Up
Cited by 43Ethics and Social Impacts of AIDigitalization, Law, and RegulationLegal and Labor Studies

Null Compliance: NYC Local Law 144 and the challenges of algorithm accountability

Open access
Wright, Lucas, Muenster, Roxana Mika, Vecchione, Briana, Qu, Tianyao, Cai, Pika (Senhuang), Smith, Alan, Comm 2450 Student Investigators, Metcalf, Jacob +1 more·2024·The 2024 ACM Conference on Fairness, Accountability, and Transparency
DOI 10.1145/3630106.3658998
Open this paper in CoChat — free account

Abstract

In July 2023, New York City became the first jurisdiction globally to mandate bias audits for commercial algorithmic systems, specifically for automated employment decisions systems (AEDTs) used in hiring and promotion. Local Law 144 (LL 144) requires AEDTs to be independently audited annually for race and gender bias, and the audit report must be publicly posted. Additionally, employers are obligated to post a transparency notice with the job listing. In this study, 155 student investigators recorded 391 employers’ compliance with LL 144 and the user experience for prospective job applicants. Among these employers, 18 posted audit reports and 13 posted transparency notices. These rates could potentially be explained by a significant limitation in the accountability mechanisms enacted by LL 144. Since the law grants employers substantial discretion over whether their system is in scope of the law, a null result cannot be said to indicate non-compliance, a condition we call "null compliance." Employer discretion may also explain our finding that nearly all audits reported an impact factor over 0.8, a rule of thumb often used in employment discrimination cases. We also find that the benefit of LL 144 to ordinary job seekers is limited due to shortcomings in accessibility and usability. Our findings offer important lessons for policy-makers as they consider regulating algorithmic systems, particularly the degree of discretion to grant to regulated parties and the limitations of relying on transparency and end-user accountability.

AI snapshot

Written from the paper’s full text.

TL;DR

This study measured early compliance with NYC Local Law 144 by having 155 student investigators model job seekers across 391 employers, finding 18 audit reports and 13 transparency notices. It introduces “null compliance” to describe how employer discretion prevents observers from treating missing audits or notices as non-compliance.

Key findings

  • Across 391 employers, researchers found 18 audit reports (5%) and 13 transparency notices (3%); among 267 employers listing NYC jobs, they found 14 audit reports (5%) and 12 transparency notices (4%), and only 11 employers published both.
  • Of 26 employer survey responses, 23 reported LL 144 did not apply to them and 3 reported it applied in some way; 2 of those 3 supplied a document meeting the authors’ definition of a bias audit, and follow-up yielded two additional audit reports.
  • Among 13 unique audit reports and 386 impact ratios, nine impact ratios were below 0.8 while all others were above it; employers omitted 20% of race/ethnicity impact ratios because subgroups represented less than 2% of the data, and one report provided no race/ethnicity impact ratios.
  • Investigators clicked through 17 pages on average for audit reports and 19 pages for transparency notices, often failing to locate them within 30 minutes; of records thought to be audit reports, 31% were audit reports and 20% were transparency notices as defined by LL 144.

Methods

  • The authors recruited 155 student investigators who searched employer websites, LinkedIn, and Indeed for no more than 30 minutes per employer, modeling job-seeker behavior without submitting applications; data collection occurred in two stages from October 24 to November 9, 2023.
  • The employer sample began with 568 employers that had hired 2021 and 2022 graduates, was reduced to 511 after exclusions, and the final list of 391 employers was a random sample from that list.
  • Investigators’ findings were verified to remove false positives; the authors counted each employer providing an audit report, included each unique audit report once, counted impact ratios separately, and surveyed employers through Consumer Reports, receiving 26 responses and two additional audit reports.

Limitations

  • The paper states the sample is limited to employers that hire Cornell students, which tend to be credentialed/professional roles, and that AEDT usage and transparency may differ in other job-market sectors; it also states the method could not observe transparency notices supplied only after someone applies.
  • The paper states that null compliance prevents determining non-compliance, makes the denominator unknown, and blocks claims about the percentage of employers in compliance or about general trends in NYC AEDT systems.
  • Evident from the text: the 30-minute search limit and reliance on student investigators may produce false negatives, and the employer survey had only 26 responses out of 387 follow-up emails, limiting what can be inferred from employer explanations.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3630106.3658998

Cites

53 of 53 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

13 of 61 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured

  • Premisecitation
    § 4.2 Notice and Consent in Other Areas of Regulation · p. 8

    Such statements can also increase discrimination when they convince minority job-applicants to disclose information that subsequently activates biased hiring decisions.

    “Such statements can also increase discrimination when they convince minority job-applicants to disclose information that subsequently activates biased hiring decisions [33].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [33] [33] Sonia K. Kang, Katherine A. DeCelles, András Tilcsik, and Sora Jun. 2016. Whitened Résumés: Race and Self-Presentation in the Labor Market. Administrative Science Quarterly 61, 3 (Sept. 2016), 469-502. https://doi.org/10.1177/0001839216639577 Publisher: SAGE Publications Inc. - source
  • Findingown data
    § 6.2 The User Experience of Compliance with LL 144 · p. 12

    Only three transparency notices mentioned the types of data collected and used in automated decision-making.

    “Only three transparency notices mentioned the types of data collected and used in automated decision-making.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 5.2 Data Verification and Aggregation · p. 10

    Across 391 employers, researchers found 18 audit reports and 13 transparency notices.

    “Across 391 employers, researchers found 18 audit reports and 13 transparency notices”
    Quote in the paperSubstantial claim
  • Findingown data
    § 6.1 Summarizing Observed Compliance with Local Law 144 · p. 10

    Of the 3 employers, 2 supplied a document that met the authors' definition of a bias audit.

    “Of the 3 employers, 2 supplied a document that met our definition of a bias audit.”
    Quote in the paperNot a substantial claim
  • Findingown data
    § 6.1 Summarizing Observed Compliance with Local Law 144 · p. 10

    Among 267 employers posting open jobs in NYC, the authors found 14 audit reports and 12 transparency notices.

    “Among 267 employers posting open jobs in NYC, we found 14 audit reports and 12 transparency notices (Table 1).”
    Quote in the paperSubstantial claim
  • May combine two claims. 41
  • Classified as probably not a substantial claim. 54

  • May combine two claims.

    Premisecitation
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 4

    The four-fifths rule is not genuinely a rule or law; it is best described as a rule of thumb or convention that is ingrained in hiring practices, and employers and developers are highly incentivized to reduce potential scrutiny by targeting that threshold.

    “The four-fifths rule is not genuinely a rule or law; it is best described as a rule of thumb or convention that is ingrained in hiring practices, and employers and developers are highly incentivized to reduce potential scrutiny by targeting that threshold.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [3] [3] Philip Bobko and Philip L Roth. 2004. THE FOUR-FIFTHS RULE FOR ASSESSING ADVERSE IMPACT: AN ARITHMETIC, INTUITIVE, AND LOGICAL ANALYSIS OF THE RULE AND IMPLICATIONS FOR FUTURE RESEARCH AND PRACTICE. In Research in Personnel and Human Resources Management . Research in Personnel and Human Resources Management, Vol. 23. Emerald Group Publishing Limited, 177-198. https: //doi.org/10.1016/S0742-7301(04)23004-3 - source
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

    • cites [69] [69] Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. 2022. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. https://doi.org/10.48550/arXiv.2202.09519 arXiv:2202.09519 [cs]. - source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4.2 Notice and Consent in Other Areas of Regulation · p. 8

    In employment law, behavioral scientists have found that equal opportunity and diversity statements can reduce applications from minority job seekers.

    “In employment law, behavioral scientists have found that equal opportunity and diversity statements can reduce applications from minority job seekers [37].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [37] [37] Andreas Leibbrandt and John A. List. 2018. Do equal employment opportunity statements backfire? Evidence from a natural field experiment on job-entry decisions . Technical Report. National Bureau of Economic Research. https://www.nber.org/papers/w25035 - source

      The Introduction reports that EEO statements in job advertisements discouraged applications from racial minorities, especially educated applicants and those in cities with white-majority populations. The paper does not establish the same result for diversity statements; its Introduction instead mentions another study finding that an explicit interest in diversity increased interest among racial minorities. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 4

    The four-fifths rule establishes a baseline against which impermissible discriminatory outcomes can be identified by comparing relative rates of selection between demographic groups.

    “The four-fifths rule establishes a baseline against which impermissible discriminatory outcomes (aka 'disparate impact') can be identified by comparing relative rates of selection between demographic groups.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [3] [3] Philip Bobko and Philip L Roth. 2004. THE FOUR-FIFTHS RULE FOR ASSESSING ADVERSE IMPACT: AN ARITHMETIC, INTUITIVE, AND LOGICAL ANALYSIS OF THE RULE AND IMPLICATIONS FOR FUTURE RESEARCH AND PRACTICE. In Research in Personnel and Human Resources Management . Research in Personnel and Human Resources Management, Vol. 23. Emerald Group Publishing Limited, 177-198. https: //doi.org/10.1016/S0742-7301(04)23004-3 - source
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

    • cites [69] [69] Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. 2022. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. https://doi.org/10.48550/arXiv.2202.09519 arXiv:2202.09519 [cs]. - source

      Appendix A.2 describes comparing demographic groups’ selection rates against the highest rate, so the rule does provide a screening threshold. But it calls the rule a rule of thumb, not a legal definition; the section on selecting a relevant comparison population says it is neither necessary nor sufficient for a court finding of disparate impact. The claim overstates what the comparison can establish by saying it identifies impermissible discriminatory outcomes. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation
    § 4.2 Notice and Consent in Other Areas of Regulation · p. 8

    Research in behavioral science and computing has found that, at best, disclosure and consent policies only help some people under some circumstances and that they can backfire.

    “Research in behavioral science and computing has found that, at best, disclosure and consent policies only help some people under some circumstances and that they can backfire [30, 52].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [30] [30] Iina Ikonen, Francesca Sotgiu, Aylin Aydinli, and Peeter W. J. Verlegh. 2020. Consumer effects of front-of-package nutrition labeling: an interdisciplinary meta-analysis. Journal of the Academy of Marketing Science 48, 3 (May 2020), 360-383. https://doi.org/10.1007/s11747-019-00663-9 - source

      The Key findings and Implications sections report that front-of-package nutrition labels have limited effects on healthier choices and can create misleading health halos or increase sales of unhealthy products. This supports the claim for one kind of information disclosure, but the paper does not address consent policies or establish the claim across behavioral science and computing generally. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [52] [52] Mostafa Purmehdi, Renaud Legoux, François Carrillat, and Sylvain Senecal. 2017. The Effectiveness of Warning Labels for Consumers: A MetaAnalytic Investigation into Their Underlying Process and Contingencies. Journal of Public Policy & Marketing 36, 1 (April 2017), 36-53. https: //doi.org/10.1509/jppm.14.047 Publisher: SAGE Publications Inc. - source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 6.3 Impact Ratios in Published Bias Audit Reports · p. 13

    In a companion study, audit industry workers indicated in interviews that, in their experience, many, if not the majority, of AEDTs on the market violate the four-fifths rule.

    “In a companion study to this one [28], audit industry workers indicated in interviews that, in their experience, many, if not the majority, of AEDTs on the market violate the four-fifths rule.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [28] [28] Lara Groves, Jacob Metcalf, Alayna Kennedy, Briana Vecchione, and Andrew Strait. 2024. Auditing Work: Exploring the New York City algorithmic bias audit regime. ACM Conference on Fairness, Accountability and Transparency (2024). https://doi.org/10.1145/3630106.3658959 - source

      Section 3.2 (p. 6) explains the four-fifths rule, and section 4.1 (p. 8) reports auditors’ interview opinions about the law. Neither reports interviewees saying that many or most AEDTs on the market violate the rule. Searches for that estimate and for tools failing the four-fifths threshold found no supporting passage in our copy. (LLM, conf=0.72)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 4

    If the less-selected group's selection rate falls below four-fifths (80%) of the baseline then there is a presumption that disparate impact may be occurring and warrants further investigation, typically through a complaint to the EEOC.

    “If the less-selected group's selection rate falls below four-fifths (80%) of the baseline then there is a presumption that disparate impact may be occurring and warrants further investigation, typically through a complaint to the EEOC.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [3] [3] Philip Bobko and Philip L Roth. 2004. THE FOUR-FIFTHS RULE FOR ASSESSING ADVERSE IMPACT: AN ARITHMETIC, INTUITIVE, AND LOGICAL ANALYSIS OF THE RULE AND IMPLICATIONS FOR FUTURE RESEARCH AND PRACTICE. In Research in Personnel and Human Resources Management . Research in Personnel and Human Resources Management, Vol. 23. Emerald Group Publishing Limited, 177-198. https: //doi.org/10.1016/S0742-7301(04)23004-3 - source
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

    • cites [69] [69] Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. 2022. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. https://doi.org/10.48550/arXiv.2202.09519 arXiv:2202.09519 [cs]. - source

      The appendix’s quotation of the Uniform Guidelines (§1607.4(D); page not specified) says a selection rate below four-fifths of the highest group’s rate is generally regarded as evidence of adverse impact. Section 2.3 describes the rule as a trigger for further regulatory scrutiny. Neither passage says that investigation typically proceeds through a complaint to the EEOC. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4.3 Characteristics of Successful Notice and Consent · p. 8

    These dark patterns are especially common in areas where firms and the public have competing incentives, including price discrimination and regulatory compliance.

    “These dark patterns are especially common in areas where firms and the public have competing incentives, including price discrimination and regulatory compliance [40].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [40] [40] Arunesh Mathur, Gunes Acar, Michael J. Friedman, Eli Lucherini, Jonathan Mayer, Marshini Chetty, and Arvind Narayanan. 2019. Dark Patterns at Scale: Findings from a Crawl of 11K Shopping Websites. Proceedings of the ACM on Human-Computer Interaction 3, CSCW (Nov. 2019), 1-32. https://doi.org/10.1145/3359183 - source

      Section 6.2 (pp. 25–26) reports dark patterns used to influence consumers for retailers’ profit and discusses practices that may violate consumer-protection rules. It does not identify price discrimination as an example or establish that dark patterns are especially common where firms’ and the public’s incentives compete. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    As critics have noted, the four-fifths rule is a fundamentally arbitrary threshold not supported by robust empirical evidence.

    “As critics have noted [69], the four-fifths rule is a fundamentally arbitrary threshold not supported by robust empirical evidence.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [69] [69] Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. 2022. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. https://doi.org/10.48550/arXiv.2202.09519 arXiv:2202.09519 [cs]. - source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 6.3 Impact Ratios in Published Bias Audit Reports · p. 13

    Multiple audit service providers further indicated that they had clients who paid for an audit for an in-scope system but declined to post it publicly.

    “Multiple audit service providers further indicated that they had clients who paid for an audit for an in-scope system but declined to post it publicly.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [28] [28] Lara Groves, Jacob Metcalf, Alayna Kennedy, Briana Vecchione, and Andrew Strait. 2024. Auditing Work: Exploring the New York City algorithmic bias audit regime. ACM Conference on Fairness, Accountability and Transparency (2024). https://doi.org/10.1145/3630106.3658959 - source

      Section 3.4 (p. 7) says an audit may have been conducted without its report being published. It does not say that multiple audit providers reported clients paying for audits of in-scope systems and then declining to publish them. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation
    § 8 CONCLUSION · p. 15

    Algorithmic employment tools are notoriously unreliable and ill-defined, imposing significant harm on workers.

    “Algorithmic employment tools are notoriously unreliable and ill-defined, imposing significant harm on workers [46, 53, 54].”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [46] [46] Arvind Narayanan. 2019. How to recognize AI snake oil. Arthur Miller Lecture on Science and Ethics (2019). https://www.cs.princeton.edu/~arvindn/ talks/MIT-STS-AI-snakeoil.pdf Publisher: Massachusetts Institute of Technology. -

      Not checked in the free run.

    • cites [53] [53] Manish Raghavan, Solon Barocas, Jon Kleinberg, and Karen Levy. 2020. Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency . 469-481. https://doi.org/10.1145/3351095.3372828 arXiv:1906.09208 [cs]. - source
    • cites [54] [54] Inioluwa Deborah Raji, I. Elizabeth Kumar, Aaron Horowitz, and Andrew Selbst. 2022. The fallacy of AI functionality. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency . 959-972. https://dl.acm.org/doi/abs/10.1145/3531146.3533158 - source

      The Introduction reports that teacher-assessment tools can mark star instructors for firing, and §3 says audits of AI hiring tools rarely mention product validation. These passages support concerns about employment-related tools, but the paper does not establish that algorithmic employment tools as a class are notoriously unreliable and ill-defined or that they generally impose significant harm on workers. Page numbers were not available in our copy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    The four-fifths rule is a general guideline that an impact ratio greater than 0.8 is considered de facto evidence (but not de jure evidence) of non-discriminatory outcomes.

    “The 'four-fifths rule' is a general guideline that an impact ratio greater than 0.8 is considered de facto evidence (but not de jure evidence) of non-discriminatory outcomes.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [3] [3] Philip Bobko and Philip L Roth. 2004. THE FOUR-FIFTHS RULE FOR ASSESSING ADVERSE IMPACT: AN ARITHMETIC, INTUITIVE, AND LOGICAL ANALYSIS OF THE RULE AND IMPLICATIONS FOR FUTURE RESEARCH AND PRACTICE. In Research in Personnel and Human Resources Management . Research in Personnel and Human Resources Management, Vol. 23. Emerald Group Publishing Limited, 177-198. https: //doi.org/10.1016/S0742-7301(04)23004-3 - source
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

    • cites [69] [69] Elizabeth Anne Watkins, Michael McKenna, and Jiahao Chen. 2022. The four-fifths rule is not disparate impact: a woeful tale of epistemic trespassing in algorithmic fairness. https://doi.org/10.48550/arXiv.2202.09519 arXiv:2202.09519 [cs]. - source

      The quoted regulation in Appendix A says that federal agencies generally do not regard an impact ratio above 0.8 as evidence of adverse impact, and Section 2.3 describes the rule as a practical screening guideline rather than a legal definition. But not treating a ratio as evidence of adverse impact is not affirmative evidence of non-discriminatory outcomes. Section 2.3 also says the rule has not generally been extended beyond employment decisions. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § 8 CONCLUSION · p. 15

    Policy experts have argued that the chief outcomes of disclosure-based regulations are the capacities they build to achieve at least some compliance and facilitate grass-roots contestation.

    “Policy experts have argued that the chief outcomes of disclosure-based regulations are the capacities they build to achieve at least some compliance [58] and facilitate grass-roots contestation [44].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [58] [58] Andrew D Selbst. 2021. An Institutional View of Algorithmic Impact Assessments. Harv. J.L. & Tech. 35 (2021), 78. https://ssrn.com/abstract=3867634 -

      Not checked in the free run.

    • cites [44] [44] Jacob Metcalf, Ranjit Singh, Emanuel Moss, Emnet Tafesse, and Elizabeth Anne Watkins. 2023. Taking Algorithms to Courts: A Relational Approach to Algorithmic Accountability. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT '23) . Association for Computing Machinery, New York, NY, USA, 1450-1462. https://doi.org/10.1145/3593013.3594092 - source

  • May combine two claims.

    Findingown data
    § 6.2 The User Experience of Compliance with LL 144 · p. 11

    The study identified 18 employer-provided bias audit reports, with some providing the same report as their AEDT vendor, leaving 13 unique reports.

    “Our study identified 18 employer-provided bias audit reports, with some providing the same report as their AEDT vendor, leaving us with 13 unique reports.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Audit reports cannot and should not be the same between employers.
    • Rests on an assumption: The final rule permits employers to pool data from multiple employers' data that use the same vendors' model, as long as the employer also contributes their data to the pooled dataset.

  • May combine two claims.

    Findingown data
    § 6.3 Impact Ratios in Published Bias Audit Reports · p. 12

    Out of 386 total measurements, nine impact ratios were below 0.8, while all others reported impact ratios above this threshold.

    “Out of 386 total measurements, nine impact ratios were below 0.8, while all others reported impact ratios above this threshold.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 6.2 The User Experience of Compliance with LL 144 · p. 11

    Of records that student investigators thought might be audit reports, 31% were audit reports and 20% were transparency notices as defined by LL 144.

    “Of records that student investigators thought might be audit reports 31% were audit reports and 20% were transparency notices as defined by LL 144.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data
    § 5 METHODS · p. 8

    By one month after the law came into effect, the authors had found 19 audit reports, 14 from employers plus 5 from software vendors not covered by LL 144.

    “By one month after the law came into effect, we had found 19 audit reports, 14 from employers plus 5 from software vendors not covered by LL 144.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability · p. 1

    Among 391 employers recorded by 155 student investigators, 18 posted audit reports and 13 posted transparency notices.

    “Among these employers, 18 posted audit reports and 13 posted transparency notices.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 6.1 Summarizing Observed Compliance with Local Law 144 · p. 10

    Of the 26 firms that responded to the survey, 23 reported that LL 144 did not apply to them and 3 reported that LL 114 applied to them in some other way.

    “Of the 26 firms that responded to our survey, 23 reported that LL 144 did not apply to them and 3 reported that LL 114 applied to them in some other way.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given
    § 7.2 Confusing User Experiences and Partial Reporting Requirements Undermine job seeker Agency · p. 14

    Transparency notices were hard to find and buried amidst many other disclosures.

    “Transparency notices were hard to find and buried amidst many other disclosures.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § A.3 Survey Response Rates · p. 22

    In total, this follow-up process yielded two audit reports that we had not previously observed.

    “In total, this follow-up process yielded two audit reports that we had not previously observed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 5

    The most relevant statistical criterion for judging discrimination in employment is consequential to the first algorithm auditing law primarily by its absence.

    “In other words, the most relevant statistical criterion for judging discrimination in employment is consequential to the first algorithm auditing law primarily by its absence .”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning
    § 5.2 Data Verification and Aggregation · p. 9

    The authors reviewed the evidence submitted with results, checked the website if necessary, and removed false positives.

    “We reviewed the evidence they submitted with their results, checked the website if necessary, and removed false positives.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 2.2 Sources of Discretion Offered to Employers · p. 3

    LL 144 does not bar systems that cause disparate impact; it only requires reporting the rates.

    “Third , the law does not bar systems that cause disparate impact [49]-it only requires reporting the rates.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [49] [49] New York City Department of Consumer and Worker Protections. 2023. Automated Employment Decision Tools (Updated). https://rules. cityofnewyork.us/rule/automated-employment-decision-tools-updated/ -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given
    § 5.1 Modeling job seekers on Employer Websites · p. 9

    In the first stage, each investigator was randomly assigned one employer to research, assigning each company to two investigators.

    “In the first stage, each investigator was randomly assigned one employer to research, assigning each company to two investigators.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 2.2 Sources of Discretion Offered to Employers · p. 4

    LL 144 creates a tension between discretion and transparency that ultimately blunts the effectiveness of the law, a result named 'null compliance'.

    “While LL 144 is not outside of typical parameters of regulatee discretion, it does appear to create a tension between discretion and transparency that ultimately blunts the effectiveness of the law, a result we name ' null compliance '.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation
    § 4.3 Characteristics of Successful Notice and Consent · p. 8

    At minimum, an informed decision is one in which people are able to reason about the possible consequences based on reliable information about the possible choices.

    “At minimum, an informed decision is one in which people are able to reason about the possible consequences based on reliable information about the possible choices [71].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [71] [71] John T. Wilbanks. 2020. Electronic Informed Consent in Mobile Applications Research. The Journal of Law, Medicine & Ethics: A Journal of the American Society of Law, Medicine & Ethics 48, 1\_suppl (March 2020), 147-153. https://doi.org/10.1177/1073110520917040 - source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4.2 Notice and Consent in Other Areas of Regulation · p. 8

    Computer scientists have observed that this model requires hundreds of hours per year per person.

    “Computer scientists have observed that this model requires hundreds of hours per year per person [42].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [42] [42] Aleecia M. McDonald and Lorrie Faith Cranor. 2008. The cost of reading privacy policies. Isjlp 4 (2008), 543. https://heinonline.org/hol-cgibin/get\_pdf.cgi?handle=hein.journals/isjlpsoc4&section=27 Publisher: HeinOnline. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisenone given
    § 4.3 Characteristics of Successful Notice and Consent · p. 8

    Notice and consent depends on informed decision-making, which relies on the accessibility, comprehensibility, and usability of information.

    “Notice and consent depends on informed decision-making, which relies on the accessibility, comprehensibility, and usability of information.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3 THE MEANING OF NULL COMPLIANCE: NOT THE SAME AS NON-COMPLIANCE · p. 5

    The law places accountability only upon the end-user of AEDT systems (the employers) to procure and post an audit, grants them near-total discretion over whether their system is in scope, and offers them many chances to move out of scope.

    “The law places accountability only upon the end-user of AEDT systems (the employers) to procure and post an audit, grants them near-total discretion over whether their system is in scope, and offers them many chances to move out of scope [48].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] [48] New York City Department of Consumer and Worker Protections. 2023. Automated Employment Decision Tools: Frequently Asked Questions. https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation
    § 5 METHODS · p. 8

    The pilot dataset was largely identical to a crowd-sourced dataset later assembled by civil society actors.

    “This pilot dataset was largely identical to a crowd-sourced dataset later assembled by civil society actors [25].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] [25] Marissa Gerchick and Brooke Watson. 2023. Tracking Automated Employment Decision Tool Bias Audits. https://github.com/aclu-national/trackingll144-bias-audits original-date: 2023-08-09T19:46:06Z. -

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 1 INTRODUCTION · p. 2

    Research can measure a compliance rate, but not a non-compliance rate under LL 144.

    “Research can measure a compliance rate, but not a non-compliance rate.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    In the case of this study, we are unable to determine whether mismeasurement is occurring.

    “In the case of this study, we are unable to determine whether mismeasurement is occurring.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationcitation
    § 2.2 Sources of Discretion Offered to Employers · p. 4

    The DCWP's definition of AEDT is an awkward mix of technical and organizational features that introduces interpretational complexity and grants employers wide latitude over whether their systems are in scope.

    “This is an awkward mix of technical features and organizational features that differs from technical definitions of machine learning [1], introducing interpretational complexity and granting employers wide latitude over whether their systems are in scope.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] [1] Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023. Fairness and machine learning: Limitations and opportunities . MIT Press. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 5

    LL 144 requires transparent auditing so job seekers are empowered to make informed choices and employers and vendors are incentivized to deploy fair(er) models, but it does not set a floor for permissible rates of discrimination.

    “LL 144 requires transparent auditing so job seekers are empowered to make informed choices and employers and vendors are incentivized to deploy fair(er) models, but it does not set a floor for permissible rates of discrimination.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [5] [5] Paul Cairney. 2016. The politics of evidence-based policy making . Springer. - source
    • cites [6] [6] Paul Cairney. 2019. Understanding public policy: theories and issues . Vol. 2. Bloomsbury Publishing. -

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § 3 THE MEANING OF NULL COMPLIANCE: NOT THE SAME AS NON-COMPLIANCE · p. 6

    A finding of null compliance in our report may occur due to any combination of conditions (1)-(7) made possible by LL 144's accountability arrangements.

    “A finding of a null compliance in our report may occur due to any combination of the following conditions made possible by LL 144's particular accountability arrangements:”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationreasoning
    § 2.2 Sources of Discretion Offered to Employers · p. 4

    Given those conditions, it is reasonable to expect that many employers and their legal advisors would be either confused about, or seek ways to legally avoid, the scope of the law.

    “Given those conditions, it is reasonable to expect that many employers and their legal advisors would be either confused about, or seek ways to legally avoid, the scope of the law.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 4.2 Notice and Consent in Other Areas of Regulation · p. 8

    The disclose-and-opt-out model of LL 144 is similar to many data privacy regulations.

    “The disclose-and-opt-out model of LL 144 is also similar to many data privacy regulations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    Since companies have discretion over whether their hiring software is governed by the law, some companies might use their discretion under LL 144 to avoid EEOC investigation.

    “Since companies have discretion over whether their hiring software is governed by the law, some companies might use their discretion under LL 144 to avoid EEOC investigation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.1 Measuring Compliance with Transparency Regulations · p. 6

    For researchers studying LL 144, deliberate ignorance could lead to under-counts in base rates of AEDT usage as well as biased estimates of racial and gender disparities in AEDTs.

    “For researchers studying LL 144, deliberate ignorance could lead to under-counts in base rates of AEDT usage as well as biased estimates of racial and gender disparities in AEDTs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 5

    This situation is central to understanding the conundrum of null compliance: one regulator (the DCWP) has demanded transparency about an activity that another regulator (the EEOC) has jurisdiction over but would not be able to observe in the usual course of business.

    “This situation is central to understanding the conundrum of null compliance: one regulator (the DCWP) has demanded transparency about an activity that another regulator (the EEOC) has jurisdiction over but would not be able to observe in the usual course of business.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 2.2 Sources of Discretion Offered to Employers · p. 3

    LL 144 directly imposes obligations upon employers only, although most employers that use AEDTs lease them from vendors and/or recruitment platforms that train and maintain the models.

    “First , LL 144 directly imposes obligations upon employers only , although most employers that use AEDTs lease them from vendors and/or recruitment platforms that train and maintain the models [49].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [49] [49] New York City Department of Consumer and Worker Protections. 2023. Automated Employment Decision Tools (Updated). https://rules. cityofnewyork.us/rule/automated-employment-decision-tools-updated/ -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 3 THE MEANING OF NULL COMPLIANCE: NOT THE SAME AS NON-COMPLIANCE · p. 5

    Consequently, when investigators, regulators, and job seekers cannot find an audit report or transparency notice, they cannot call it non-compliance.

    “Consequently, when investigators, regulators, and job seekers cannot find an audit report or transparency notice, they cannot call it non-compliance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    While over-compliance might generate public goods by giving job seekers more choices and increasing the transparency of hiring software, it also undermines the accuracy of research on the use of AEDT systems and compliance with the law.

    “While over-compliance might generate public goods by giving job seekers more choices and increasing the transparency of hiring software, it also undermines the accuracy of research on the use of AEDT systems and compliance with the law.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    Since LL 144 gives job seekers the right to request a costly alternative to the AEDT, employers have strong financial incentives to make disclosures difficult to find.

    “Since LL 144 gives job seekers the right to request a costly alternative to the AEDT, employers have strong financial incentives to make disclosures difficult to find.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning
    § 5.1 Modeling job seekers on Employer Websites · p. 9

    To minimize false negatives, investigators were encouraged to record any and all material that might plausibly be related to LL 144.

    “To minimize false negatives, we encouraged investigators to record any and all material that might plausibly be related to LL 144.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability · p. 1

    LL 144 requires AEDTs used in hiring or promotion to be independently audited annually for race and gender bias, and the audit report must be publicly posted.

    “Local Law 144 (LL 144) requires AEDTs to be independently audited annually for race and gender bias, and the audit report must be publicly posted.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Appendix A.1

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2.2 Sources of Discretion Offered to Employers · p. 4

    While there is no reliable source about what percentage of employers utilize AEDTs, surveys show that AI systems for recruitment, hiring, and HR are widespread and rapidly growing in reach.

    “While there is no reliable source about what percentage of employers utilize AEDTs of any type, what surveys do exist show that AI systems for recruitment, hiring, and HR are widespread and rapidly growing in reach [31, 62].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [31] [31] Indeed. 2023. The Indeed Global AI Survey: Your Guide to the Future of Hiring . Technical Report. Indeed. https://www.indeed.com/lead/the-indeedai-report?hl=en#form -

      Not checked in the free run.

    • cites [62] [62] SHRM. 2022. Automation & AI in HR. https://advocacy.shrm.org/SHRM-2022-Automation-AI-Research.pdf?\_ga=2.112869508.1029738808. 1666019592-61357574.1655121608 -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    LL 144 includes two decision-rules that invite creative behavior.

    “LL 144 includes two decision-rules that invite creative behavior.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data
    § 5.3 Survey of Employers · p. 10

    The authors contacted employers with two rounds of emails from Consumer Reports to further verify whether the research team might have missed public information about LL 144.

    “We also contacted employers with two rounds of emails from Consumer Reports to further verify whether our research team might have missed public information about LL 144.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.1 Measuring Compliance with Transparency Regulations · p. 7

    For LL 144, access control could lead to an under-count of AEDTs, with mostly favorable audit reports published.

    “For LL 144, access control could lead to an under-count of AEDTs, with mostly favorable audit reports published.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation
    § 2.3 Local Law 144 and Anti-Discrimination Law · p. 5

    Employers complying with LL 144 by posting audit reports might also be opening themselves to other types of liability, creating conflicting regulatory pressures.

    “Thus, employers complying with LL 144 by posting audit reports might also be opening themselves to other types of liability, creating conflicting regulatory pressures [43].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [43] [43] Jacob Metcalf. 2023. What federal agencies can learn from New York City's AI hiring law. https://thehill.com/opinion/technology/4360523-whatfederal-agencies-can-learn-from-new-york-citys-ai-hiring-law/ -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § 3 THE MEANING OF NULL COMPLIANCE: NOT THE SAME AS NON-COMPLIANCE · p. 6

    Null compliance should not be taken to indicate non-compliance in this paper-absent legal investigatory powers that we lack, we cannot say that any given case of null compliance is actually non-compliance.

    “Null compliance should not be taken to indicate non-compliance in this paper-absent legal investigatory powers that we lack, we cannot say that any given case of null compliance is actually non-compliance .”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    The FAQ notes that other laws on employment discrimination still apply outside the scope of LL 144, which would presumably include the EEOC's definition of disparate impact.

    “The FAQ does note that other laws on employment discrimination still apply outside of the scope of LL 144, which would presumably include the EEOC's definition of disparate impact [14, 48, 50].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [14] [14] Equal Employment Opportunity Commission. 2023. Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964. https://www.eeoc.gov/laws/guidance/select-issues-assessingadverse-impact-software-algorithms-and-artificial -

      Not checked in the free run.

    • cites [48] [48] New York City Department of Consumer and Worker Protections. 2023. Automated Employment Decision Tools: Frequently Asked Questions. https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    This guidance is situated in the EEOC's long-standing decision tree for determining acceptable selection practices, which offers few hard lines amenable to auditing and relies heavily on judgment calls and precedents.

    “Nonetheless, this guidance is situated in the EEOC's long-standing decision tree for determining acceptable selection practices, which offers few hard lines amenable to auditing and relies heavily on judgment calls and precedents [12].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [12] [12] Equal Employment Opportunity Commission. 2007. Employment Tests and Selection Procedures. https://www.eeoc.gov/laws/guidance/employmenttests-and-selection-procedures -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    If introduced, such safe harbors would protect employers from disincentives to disclose information about discriminatory algorithms, as is the case in other areas of law.

    “If introduced, such safe harbors would protect employers from disincentives to disclose information about discriminatory algorithms, as is the case in other areas of law [68].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [68] [68] Peter V. B. Unger, Karen Van Esen, Alexander Birkhold, Mohammed T. Farooqui, and Elizabeth Satarov. 2023. DOJ Voluntary Self-Disclosures Safe Harbor Policy. (Oct. 2023). https://natlawreview.com/article/doj-announces-new-mergers-acquisitions-safe-harbor-policy-voluntary-self -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given
    § 8 CONCLUSION · p. 15

    If fewer employers use hiring algorithms, the rate of algorithmic bias will definitionally go down, even though it is hard to know if that would increase or decrease employment discrimination overall.

    “If fewer employers use hiring algorithms, the rate of algorithmic bias will definitionally go down, even though it is hard to know if that would increase or decrease employment discrimination overall.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 7.3 Could LL 144 Reduce Employment Discrimination? · p. 15

    Absent some sort of safe harbor, any employer that uses a system with an impact ratio below 0.8 will need to decide if complying with LL 144's transparency requirements will provide information for EEOC enforcement or private litigation.

    “Absent some sort of safe harbor, any employer that uses a system with an impact ratio below 0.8 will need to decide if complying with LL 144's transparency requirements will provide information for EEOC enforcement or private litigation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    Federal regulators have held hearings on the appropriateness of the four-fifths rule for AI use cases.

    “including holding hearings on the appropriateness of the four-fifths rule for AI use cases [16, 70].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [16] [16] Equal Employment Opportunity Commission. 2023. Navigating Employment Discrimination in AI and Automated Systems: A New Civil Rights Frontier - Transcript. https://www.eeoc.gov/meetings/meeting-january-31-2023-navigating-employment-discrimination-ai-and-automatedsystems-new/transcript -

      Not checked in the free run.

    • cites [70] [70] Christine Webber. 2023. AI Bias Panel Shows EEOC Should Ditch Four-Fifths Rule - Law360 Expert Analysis | Cohen Milstein. https://www. cohenmilstein.com/update/ai-bias-panel-shows-eeoc-should-ditch-four-fifths-rule-%E2%80%93-law360-expert-analysis -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.1 Key Terms for understanding employment discrimination · p. 20

    Falling below an impact ratio of 0.8 is not considered to be automatic evidence of illegal discrimination absent an investigation that considers many other factors that may legally justify a discriminatory outcome, such as legitimate relevance of the features to the job.

    “Falling below an impact ratio of 0.8 is not considered to be automatic evidence of illegal discrimination absent an investigation that considers many other factors that may legally justify a discriminatory outcome, such as legitimate relevance of the features to the job.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given
    § Ethical Review · p. 19

    We have taken measures to minimize risk, contacting employers to give them a chance to correct any errors.

    “We have nonetheless taken measures to minimize risk, contacting employers to give them a chance to correct any errors.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given
    § Ethical Review · p. 19

    We submitted this study for ethics review to Cornell University in protocol #IRB0147978.

    “We submitted this study for ethics review to Cornell University in protocol #IRB0147978.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    LL 144 makes no explicit mention of a 0.8 threshold or any other threshold.

    “LL 144 makes no explicit mention of a 0.8 threshold or any other threshold.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [49] [49] New York City Department of Consumer and Worker Protections. 2023. Automated Employment Decision Tools (Updated). https://rules. cityofnewyork.us/rule/automated-employment-decision-tools-updated/ -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data
    § A.3 Survey Response Rates · p. 21

    Of all the employers in our sample, 29 listed no email or contact form or had a contact form with a character limit that was so low as to make the message incomplete.

    “Of all the employers in our sample, 29 listed no email or contact form or had a contact form with a character limit that was so low as to make the message incomplete.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given
    § Ethical Review · p. 19

    We have de-coupled the identities of student investigators from the dataset.

    “We have also de-coupled the identities of student investigators from the dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning
    § Ethical Review · p. 19

    The IRB concluded that the study does not qualify as human subjects research regulated under the Common Rule.

    “they concluded that the study does not qualify as human subjects research regulated under the Common Rule.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data
    § A.3 Survey Response Rates · p. 21

    Of the emails we did send in the first round, 15 bounced or received a reply saying the address was not monitored.

    “Of the emails we did send in the first round, 15 bounced or received a reply saying the address was not monitored.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § A.2 Key Legal and Historical Details · p. 21

    Because LL 144 does not take a stance on permissible thresholds, it is also silent on remediation, and thereby limits the utility of the law for driving and measuring reduction of discrimination over time.

    “Because LL 144 does not take a stance on permissible thresholds, it is also silent on remediation, and thereby limits the utility of the law for driving and measuring reduction of discrimination over time.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data
    § A.3 Survey Response Rates · p. 22

    Among the 387 follow-up survey emails that we sent, 26 employers completed the survey, and 15 contacted us in some other way, either by email or phone.

    “Among the 387 follow-up survey emails that we sent, 26 employers completed the survey, and 15 contacted us in some other way, either by email or phone.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.1 Key Terms for understanding employment discrimination · p. 20

    The EEOC defines the Four-Fifths Rule in CFR §1607.4(D).

    “The EEOC defines the Four-Fifths Rule in CFR §1607.4(D) 15 .”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] [50] Code of Federal Regulations. 2024. 29 CFR Part 1607 - General Principles. https://www.ecfr.gov/current/title-29/part-1607/subject-groupECFRdb347e844acdea6 -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § A.3 Survey Response Rates · p. 21

    The first case of algorithmic hiring bias-for age discrimination-was settled with the EEOC in 2023, but it was for a system that automatically screened out all older applicants and not a typical disparate impact scenario.

    “The first case of algorithmic hiring bias-for age discrimination-was settled with the EEOC in 2023, but it was for a system that automatically screened out all older applicants and not a typical disparate impact scenario [15, 26].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [15] [15] Equal Employment Opportunity Commission. 2023. iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit. https://www.eeoc. gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit -

      Not checked in the free run.

    • cites [26] [26] Annelise Gilbert. 2023. EEOC Settles First-of-Its-Kind AI Bias in Hiring Lawsuit. Bloomberg Law (Aug. 2023). https://news.bloomberglaw.com/dailylabor-report/eeoc-settles-first-of-its-kind-ai-bias-lawsuit-for-365-000 -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § 7.1 Null Compliance in LL 144 Hinders Enforcement and Transparency · p. 14

    Within audit reports, null compliance can obscure employment disparities for minority communities due to provisions in the law that permit employers to exclude data from audit reports.

    “Within audit reports, null compliance can obscure employment disparities for minority communities due to provisions in the law that permit employers to exclude data from audit reports.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning
    § 7.3 Could LL 144 Reduce Employment Discrimination? · p. 14

    LL 144 does not set a floor for permissible impact ratios and instead relies on the four-fifths rule implicitly; it makes no reference to the rule but requires auditors to use exactly the equations and disclose results that any disparate impact investigation would make use of.

    “LL 144 does not set such a floor and instead relies on the 'four-fifths rule' implicitly; it makes no reference to the rule but requires auditors to use exactly the equations and disclose results that any disparate impact investigation would make use of.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § A.2 Key Legal and Historical Details · p. 21

    Given the absence of any safe harbors, employers complying with LL 144 by posting audit reports might also be opening themselves to other types of liability.

    “Given the absence of any safe harbors, employers complying with LL 144 by posting audit reports might also be opening themselves to other types of liability.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: No federal body has established guidance about safe harbors for audits conducted in compliance with a local or state law.

  • May combine two claims.

    Interpretationreasoning
    § 7.1 Null Compliance in LL 144 Hinders Enforcement and Transparency · p. 13

    Null compliance creates strong barriers to enforcement; investigators or complainant job seekers cannot tell whether the law does not apply, whether an employer is refusing to conduct a bias audit, or whether they are hiding an unfavorable audit.

    “Null compliance creates strong barriers to enforcement. In such cases, investigators or complainant job seekers cannot tell whether the law does not apply, whether an employer is refusing to conduct a bias audit, or whether they are hiding an unfavorable audit.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation
    § Adverse Impacts · p. 19

    We adopted best practices from journalism to ensure the accuracy and reliability of our data, giving employers the opportunity to correct any errors.

    “Since this study focuses on organizations, we adopted best practices from journalism [23] to ensure the accuracy and reliability of our data, giving employers the opportunity to correct any errors.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] [23] Gene Foreman, Daniel R. Biddle, Emilie Lounsberry, and Richard G. Jones. 2022. The ethical journalist: Making responsible decisions in the digital age . John Wiley & Sons. -

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § 7.3 Could LL 144 Reduce Employment Discrimination? · p. 15

    Because an impact ratio above 0.8 is not a guaranteed shield against litigation, publishing even apparently good results could conceivably feed adverse litigation outcomes.

    “And because an impact ratio above 0.8 is not a guaranteed shield against litigation, publishing even apparently good results could conceivably feed adverse litigation outcomes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 7.1 Null Compliance in LL 144 Hinders Enforcement and Transparency · p. 13

    Overall, the problem of null compliance hinders the law's function and prevents researchers from knowing if the law's purposes are being achieved.

    “Overall, the problem of null compliance hinders the law's function and prevents researchers from knowing if the law's purposes are being achieved.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The theory of change behind LL 144 assumes that employers will make information accessible to job seekers who can make informed decisions.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § A.2 Key Legal and Historical Details · p. 21

    The concept of disparate impact and the four-fifths rule loom large in discussions of algorithmic bias because the unintentional, systemic discrimination described by disparate impact is often considered structurally analogous to algorithmic discrimination.

    “The concept of disparate impact and the four-fifths rule loom large in discussions of algorithmic bias because the unintentional, systemic discrimination described by disparate impact is often considered structurally analogous to algorithmic discrimination [2].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2] [2] Solon Barocas and Andrew D. Selbst. 2016. Big data's disparate impact. California law review (2016), 671-732. https://www.jstor.org/stable/24758720 Publisher: JSTOR. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning
    § 7.1 Null Compliance in LL 144 Hinders Enforcement and Transparency · p. 14

    Since employers are in practice excused from reporting impact ratios for groups that they have hired the least, LL 144 is unable to achieve transparency or accountability for those groups.

    “Since employers are in practice excused from reporting impact ratios for groups that they have hired the least, LL 144 is unable to achieve transparency or accountability for those groups.”
    Quote in the paperNot a substantial claim
79 of 79

  • The paper relies on this without support.

    The outcomes of LL 144's policy components have yet to be established in real-world algorithm governance.

    “Yet the outcomes of these policy components have yet to be established in real-world algorithm governance.”

  • The paper relies on this without support.

    The DCWP establishes no central public repository for audit reports, relying solely on employers to post their own.

    “Second , the DCWP establishes no central public repository for audit reports, relying solely on employers to post their own.”

  • The paper relies on this without support.

    It does not appear that any vendors currently offer tools that fully replace human discretion in the hiring process with complete automation.

    “Although LL 144 is built with them in mind, it does not appear that any vendors currently offer tools that fully replace human discretion in the hiring process with complete automation; thus any attempt to track the effect of the law would need to focus on systems that 'substantially assist' humans.”

  • The paper relies on this without support.

    No complaints have been submitted to the DCWP at the point of publication.

    “To the best of the authors' knowledge, no complaints have been submitted to the DCWP at the point of publication.”

  • The paper relies on this without support.

    There is no reliable source about what percentage of employers utilize AEDTs of any type.

    “While there is no reliable source about what percentage of employers utilize AEDTs of any type, what surveys do exist show that AI systems for recruitment, hiring, and HR are widespread and rapidly growing in reach [31, 62].”

  • The paper relies on this without support.

    Employers have limited guidance and no case law to inform the user experience.

    “Employers have limited guidance and no case law to inform the user experience.”

  • The authors state this as an assumption.

    The theory of change behind LL 144 assumes that employers will make information accessible to job seekers who can make informed decisions.

    “The theory of change behind LL 144 assumes that employers will make information accessible to job seekers who can make informed decisions.”

  • The paper relies on this without support.

    Audit reports cannot and should not be the same between employers.

    “Yet since audit reports cannot and should not be the same between employers, they are required to publish unique reports.”

  • The paper relies on this without support.

    The final rule permits employers to pool data from multiple employers' data that use the same vendors' model, as long as the employer also contributes their data to the pooled dataset.

    “The final rule permits employers to pool data from multiple employers' data that use the same vendors' model, as long as the employer also contributes their data to the pooled dataset.”

  • The paper relies on this without support.

    No federal body has established guidance about safe harbors for audits conducted in compliance with a local or state law.

    “Nor has any federal body established guidance about safe harbors for audits conducted in compliance with a local or state law.”

  • The paper relies on this without support.

    In many algorithmic bias detection and mitigation tools, the four-fifths rule is treated as a foundational, standard measure of the line between acceptable and unacceptable discrimination, despite the fact that it is domain-specific inside of US law.

    “Indeed, in many algorithmic bias detection and mitigation tools, the four-fifths rule is treated as a foundational, standard measure of the line between acceptable and unacceptable discrimination, despite the fact that it is domain-specific inside of US law.”

  • The paper relies on this without support.

    The four-fifths rule is historically important, well-litigated, and embedded in U.S. anti-discrimination jurisprudence, and therefore familiar and unavoidable.

    “But it is historically important, well-litigated, and embedded in U.S. anti-discrimination jurisprudence, and therefore familiar and unavoidable.”

  • The paper relies on this without support.

    Falling below an impact ratio of 0.8 is a strong signal that a hiring process could receive scrutiny and requires additional justification.

    “However, falling below an impact ratio of 0.8 is a strong signal that a hiring process could receive scrutiny and requires additional justification.”
13 of 13

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Several tools surfaced in the survey (e.g., from Holistic.ai, Credo.ai) were explicitly marketed for use for NYC Local Law 144 compliance, but studies of audit practice suggest the produced measures may not be reliable or legally compatible.

    “For instance, several tools surfaced in our survey (e.g., from Holistic.ai, Credo.ai) were explicitly marketed for use for NYC Local Law 144 compliance, but studies of audit practice suggest the produced measures may not be reliable or legally compatible [30, 84, 86].”

    From Towards AI Accountability Infrastructure: Gaps and Opportunities in AI Audit Tooling (2025) · cites this paper as [84]

    Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023. Fairness and machine learning: Limitations and opportunities. MIT Press.
  2. Paul Cairney. 2019. Understanding public policy: theories and issues. Vol. 2. Bloomsbury Publishing.
  3. Lauren Kirchner. 2017. New York City Moves to Create Accountability for Algorithms. ProPublica (Dec. 2017). https://www.propublica.org/article/new-york-city-moves-to-create-accountability-for-algorithms
  4. Aleecia M. McDonald and Lorrie Faith Cranor. 2008. The cost of reading privacy policies. Isjlp 4 (2008), 543. https://heinonline.org/hol-cgi-bin/get_pdf.cgi?handle=hein.journals/isjlpsoc4&section=27 Publisher: HeinOnline.
  5. Don Norman. 2013. The design of everyday things: Revised and expanded edition. Basic books.
  6. Robert N. Proctor and Londa Schiebinger. 2008. Agnotology: The making and unmaking of ignorance. (2008). https://philarchive.org/archive/PROATM
  7. J. Richard, Judith D. Singer, John B. Willett, and Richard J. Light. 2009. By design: Planning research on higher education. Harvard University Press.
  8. Peter V. B. Unger, Karen Van Esen, Alexander Birkhold, Mohammed T. Farooqui, and Elizabeth Satarov. 2023. DOJ Voluntary Self-Disclosures Safe Harbor Policy. (Oct. 2023). https://natlawreview.com/article/doj-announces-new-mergers-acquisitions-safe-harbor-policy-voluntary-self

Cited by

Sorted by citations

Related

Authors
Wright, L., Muenster, R. M., Vecchione, B., Qu, T., Cai, P. (., Smith, A., Comm 2450 Student Investigators, Metcalf, J., & Matias, J. N.
Affiliations
  • Cornell University (US)
  • Data & Society Research Institute (US)
  • Consumer Reports, USA
ORCID
Venue
The 2024 ACM Conference on Fairness, Accountability, and Transparency
Publisher
ACM
Published
2024-06-03
Pages
1701-1713
DOI
10.1145/3630106.3658998
OpenAlex
W4399362990
Semantic Scholar
450320aefe4065ab475af17175dfe249e64fbd49
Cited by
43
References
40
Citations by year
2026: 20 · 2025: 18 · 2024: 5
FWCI
19.61
Type
conference paper
Language
en
Open access
Licence
Locations
  • ACM Conference on Fairness, Accountability, and Transparency (FAccT) (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
  • arXiv (Cornell University)
Funders
National Science Foundation (1704425)
Topics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Social Sciences › Law › Digitalization, Law, and Regulation
  • Social Sciences › Business, Management and Accounting › Industrial relations › Legal and Labor Studies
Keywords
algorithmic accountability, algorithmic discrimination, employer compliance, employer discretion, algorithmic regulation, New York City

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-30.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-30.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review