Perspectives | Opinion and legal analysis
By Sally Vazquez-Castellanos
Publication date and time: October 28, 2026, at 1:56 p.m. PDT
This article grew out of today’s discussion with ChatGPT, a generative AI tool, about automated hiring, civil rights, and the limits of New York City’s law intended to improve fairness in the use of automated tools when screening applicants. It presents my opinions alongside the public sources discussed below.
Artificial intelligence guidelines promise fairness, transparency, human oversight, and accountability. Before accepting those assurances, I want to know what they mean for the person whose application never reaches an interview.
Who selected the criteria? What does the score measure? Can the applicant discover an error? Who has authority to reconsider? And what happens when the employer cannot explain why an otherwise qualified person was excluded?
Those questions determine whether a safeguard works. A policy document, a vendor assurance, and a completed audit will never address those considerations, or any harm associated with the use of those systems. In fact, I go even further and emphatically believe that it only continues to institutionalize racism, discrimination and the abhorred class systems that divide communities in America.
My concern is that employers are delegating consequential judgments to tools whose operation applicants cannot meaningfully examine. I do not believe the present arrangements provide sufficient openness, transparency, or fairness. It has been reported that evidence to date fails to show that New York City’s audit requirement has produced a marked improvement in workforce diversity or reduced discriminatory hiring. That absence of demonstrated improvement is not proof that every tool discriminates or that the law provides no benefit. However, it is a reason to demand better evidence. It may also be a “sign” that real lawyers instead of automated tools are needed to address all of the many dispute impacts that will inevitably result from our reliance on a digital age that has failed to adequately address these types of problems in American life.
What AI guidelines actually promise
In plain terms, AI guidelines are instructions about how a system should be built, checked, and used. Some are voluntary. Others are legal requirements. Neither a promise to follow guidelines nor a completed checklist proves that a particular decision was fair.
The EU AI Act attempts to illustrate the difference. It identifies certain uses of AI in hiring as “high risk” because mistakes or discrimination can affect a person’s livelihood. That designation does not mean the government has approved the tool. It means the tool falls into a category requiring additional safeguards, subject to the Act’s scope, exceptions, and application dates.
Those safeguards address practical questions. Documentation should help establish what the system was designed to do. Data governance concerns the information used to develop and evaluate it. Logs help reconstruct what happened. Human oversight should enable a person to recognize limitations and intervene. Accuracy and robustness concern how reliably the system performs.
Consider a hypothetical tool that ranks résumés. Keeping a record of its scores is useful. It does not establish that the ranking criteria measure the job’s actual requirements. Having a recruiter approve the list is useful only if the recruiter can examine and question it. That alone should tell you a lot because you did need to be skilled at assessing the data.
An audit may show that required checks occurred. It may still leave unanswered why a qualified applicant was rejected, whether important information was wrong, or whether the criteria penalized a protected group. The safeguards matter, but they are not a certificate of fairness.
American employers also operate under existing civil-rights law. They do not acquire permission to discriminate by purchasing software. Selection procedures with racial disparities require substantive examination under the applicable legal framework, including job-relatedness, business necessity, and less discriminatory alternatives.
The practical problem is the distance between a principle and its implementation. An employer can promise human oversight while giving its recruiter little time, limited technical information, and no effective authority to question a score.
A human signature is not necessarily independent judgment. Perhaps much more oversight is needed than that signature.
What New York City requires
Local Law 144 regulates certain automated employment decision tools used to screen applicants or employees for promotion. For covered use, employers and employment agencies must ensure a recent independent bias audit, publish the required summary, and provide required notices. Enforcement began July 5, 2023.
The implementing rules reach specified reliance on outputs: sole reliance, weighting an output more heavily than any other individual criterion, or using it to override conclusions from other factors. A person making the final decision does not automatically remove the process from coverage.
But the law is narrower than a comprehensive prohibition on discriminatory employment technology. DCWP guidance excludes sourcing people who have not applied for a specific position. It also states that the law does not require particular corrective action based on audit results. Other antidiscrimination laws remain applicable. Employment attorneys are still very much in play here because I do not envision that the racial disparities, or the socio-economic disparities will improve and ordinary hard working Americans will inevitably be hurt and aged out at a time in their lives like we have never seen before.
That is the distinction employers and applicants must understand: completing an audit is not a finding that an assessment is lawful, appropriate, or accurate for a particular applicant.
What the evidence actually shows
The most useful research does not establish that the law has created a more representative workforce.
The 2024 study Null Compliance examined 391 employers and located 18 audit reports and 13 transparency notices. Its important qualification was that missing disclosures did not establish violations: researchers could not reliably determine which employers used covered tools. The study also documented difficulties locating and understanding disclosures.
That is a serious accountability problem. An applicant cannot evaluate a process that remains invisible, and an outside researcher cannot measure compliance without knowing who is subject to the requirement.
A 2025 study, Auditing the Audits, examined 116 audits contained in 44 reports collected through early November 2024. It identified incomplete demographic information, opaque aggregation, problematic test data, and metrics that could fail to reflect actual deployment. These were critiques of audit quality, not a citywide causal study of hiring outcomes.
The official enforcement record is also troubling. In December 2025, the State Comptroller reported deficiencies in DCWP’s enforcement during July 2023 through June 2025. Among other findings, only three of twelve test calls were correctly routed, and state auditors identified at least seventeen potential compliance problems in a review where DCWP had identified one. DCWP disputed some findings but accepted most recommendations wholly or partly. These were potential violations, not seventeen adjudicated discrimination findings.
Taken together, these sources support skepticism about visibility, audit quality, and enforcement. They do not demonstrate that audited tools have delivered more equitable hiring or a more diverse workforce.
The old “Look” inside a new score
Abercrombie’s historical “Look” litigation offers a useful warning.
In Samantha Elauf’s case, the appellate record described testimony that a manager instructed an interviewer to lower her appearance score because of her headscarf, bringing her below the hiring threshold. The Supreme Court subsequently rejected the appellate court’s actual-knowledge requirement for the religious-accommodation disparate-treatment claim.
The related Gonzalez litigation challenged alleged racial and gender exclusion in employment practices associated with the company’s preferred image. It resolved through settlement and monitored changes, without an admission of liability.
These were human decisions, not AI cases. They do not establish current wrongdoing by Abercrombie.
They do show why a score deserves scrutiny. If an employer’s historical definition of a successful employee reflects racial exclusion, a system trained to reproduce that judgment may reproduce the exclusion. That is a risk to investigate in a particular model, not an assumption about every algorithm.
A machine can consistently reproduce a criterion that should never have governed access to employment.
The application can become the barrier
A job seeker may need to submit a résumé, complete an assessment, or supply other information simply to be considered. When an outside platform becomes the only practical route, declining its process may mean losing the opportunity.
That does not automatically make the collection unlawful. It makes the terms consequential.
What is necessary for the application? What information is retained? Is it reused for product improvement or model development? Can an applicant correct an inference without disappearing from consideration?
LinkedIn’s public documentation describes AI hiring agents using profile information, available résumés, screening answers, and preferences. It also describes employers supplementing that information with recruiting notes and application material. Those disclosures do not establish what any particular employer received or used.
Disney, as an illustrative employer discussed in this inquiry, publicly describes AI assistance in candidate discovery and skills summaries while stating that people make advancement and hiring decisions. That statement alone neither proves discrimination nor resolves coverage under my Local Law 144 hypothetical. The actual process remains the relevant subject.
The same discipline applies to an applicant’s X profile: its public existence does not establish that a recruiter reviewed it, that a vendor scored it, or that it affected rejection.
The evidence must connect collection, interpretation, and decision.
Diversity cannot be reduced to a quota or a dashboard
I want an American workforce that reflects the breadth of American talent. I do not mean that every employer must mechanically reproduce national population percentages or select individuals according to racial quotas.
I mean that access should not depend on inherited advantage disguised as “fit,” irrelevant proxies, inaccessible assessments, or criteria that fail to measure the work.
Equal opportunity requires examining who enters the applicant pool, who completes the process, who advances, and why. A favorable aggregate number may conceal a barrier at one stage. A disparity requires careful investigation rather than an automatic legal conclusion.
This work does not require a political label. Fair Hiring and Employment Accountability describes it plainly: justified criteria, accessible processes, independent scrutiny, and responsibility for correction.
What accountability would require
Correction should focus on removing unlawful barriers. It should not mean using different racial passing scores or manipulating an individual’s result because of race; Title VII expressly prohibits race-norming employment tests.
Employers should assess tools in the setting where they are actually used. That means examining the role, applicant population, cutoffs, overrides, and changes over time. Vendor documentation should support that examination, not replace it.
Applicants should receive understandable information and a meaningful way to correct errors. Reviewers need enough information and authority to reconsider outcomes. Relevant records must be preserved so a lawyer, investigator, or court can reconstruct the decision.
New York City should strengthen proactive enforcement, technical review, and coordination between audit-compliance officials and civil-rights investigators. A centralized registry, stronger audit standards, and requirements for explaining and addressing material disparities deserve consideration. These are proposed improvements, not descriptions of existing duties.
The lesson extends beyond hiring. A facial-recognition candidate, a vehicle-camera observation, and an employment score are different outputs governed by different rules. None should quietly become an authoritative conclusion about a person without appropriate verification and accountability.
Nor do Meta, Disney, universities, and technology vendors have identical duties merely because their operations reach large populations. Their common governance question is whether the people affected can obtain answers, challenge mistakes, and secure effective correction.
For lawyers, the work is far from over. It includes examining what a system measures, tracing the decision, testing the institution’s explanation, and connecting evidence to substantive rights.
My objection is to exclusion acquiring the appearance of technical legitimacy. Historical racial hierarchies can be reproduced through contemporary data and institutional choices. Whether that occurred in a particular case requires proof. The possibility is too serious to answer with an audit certificate.
Connect audit findings to civil rights enforcement
I would strengthen this law by connecting the audit to a documented correction process and to the agencies that enforce substantive civil rights. Discovering a disparity should begin an inquiry rather than end with publication of a table.
Existing federal law supplies an important framework. Under Title VII, a claimant challenging disparate impact generally must identify a particular employment practice and establish that it causes a significant adverse impact on a protected group. The employer then must demonstrate that the practice is job-related for the position and consistent with business necessity. Even if it does, an available less discriminatory alternative that meets the business need can remain legally significant if the employer refuses to adopt it. This is different from proving intentional discrimination.
In New York, the State and City Human Rights Laws supply additional protections, with their own standards, coverage, procedures, and remedies. Those standards should not be treated as identical to federal law. DCWP handles Local Law 144 audit-and-notice compliance; its guidance directs discrimination complaints to the NYC Commission on Human Rights. An audit violation does not automatically establish discrimination, and audit compliance does not defeat a civil-rights claim.
California offers a further example of linking technology to existing rights. Its employment regulations, effective October 1, 2025, clarify how antidiscrimination protections apply to automated-decision systems and address preservation of employment records, including automated-decision data. They do not establish that every AI-assisted decision is unlawful.
Correction should begin with the actual barrier: the assessment, cutoff, job requirements, applicant population, and stage at which people are excluded. A statistical warning should prompt investigation. It should not be treated as an automatic finding of liability.
The employer should document what it found, the evidence supporting the challenged criterion, and the steps it will take. Someone must own that work, with a timetable for changes and follow-up testing. If an effective alternative would have less discriminatory impact, it deserves serious consideration. Where a serious unexplained problem remains, I would consider pausing the affected screening function and using a documented alternative while the investigation proceeds.
Accountability should also reach the people already affected. Applicants need a workable way to correct inaccurate information and obtain appropriate reconsideration. Whether compensation or other relief is warranted depends on the facts and applicable law; an audit does not automatically award a remedy.
Enforcement agencies should coordinate referrals and technical review within their authority and confidentiality obligations. DCWP, the NYC Commission on Human Rights, the State Division of Human Rights, and the EEOC have different responsibilities. Applicants should receive clear information about the proper forum and filing deadlines rather than be left to navigate those distinctions alone.
Finally, the employer should test whether the correction actually removed the unjustified barrier. A rewritten policy is not evidence of success unless subsequent decisions support that conclusion.
These are recommendations for employer practice and regulatory improvement. Some require new rules or legislation; others involve better use of existing authority. Local Law 144 does not currently mandate this entire sequence.
Employers should also require vendor cooperation, access to relevant records, and notice of material model changes. An agreement assigning work to a supplier does not erase the employer’s own obligations.
For me, accountability means that a person can identify the challenged decision, obtain meaningful review through the appropriate process, and secure correction or relief where the law warrants it. Audit compliance and civil-rights enforcement must work together.
New York City has created a mechanism for demanding some information. It has not established that automated hiring is fair.
The employer still owes an answer for the decision.
Sources and legal authorities
Regulation (EU) 2024/1689, arts. 6, 9–15 & annex III. European Commission, AI Act, https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai.
EEOC, Section 15 Race and Color Discrimination, https://www.eeoc.gov/laws/guidance/section-15-race-and-color-discrimination; EEOC, Questions and Answers to Clarify and Provide a Common Interpretation of the Uniform Guidelines on Employee Selection Procedures, questions 38–39, https://www.eeoc.gov/es/node/130157. See also 42 U.S.C. § 2000e-2(k); Griggs v. Duke Power Co., 401 U.S. 424 (1971).
N.Y.C. Admin. Code §§ 20-870–20-874; NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools, https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page.
6 R.C.N.Y. § 5-300, https://codelibrary.amlegal.com/codes/newyorkcity/latest/NYCrules/0-0-0-138393.
DCWP, Automated Employment Decision Tools Frequently Asked Questions, parts I and II (June 29, 2023), https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf.
Lucas Wright et al., Null Compliance NYC Local Law 144 and the Challenges of Algorithm Accountability (2024), https://arxiv.org/abs/2406.01399.
Marissa Kumar Gerchick et al., Auditing the Audits Lessons for Algorithmic Accountability from Local Law 144’s Bias Audits (2025), https://doi.org/10.1145/3715275.3732004; public paper: https://facctconference.org/static/docs/facct2025-206archivalpdfs/facct2025-final14-acmpaginated.pdf.
Office of the New York State Comptroller, Enforcement of Local Law 144 Automated Employment Decision Tools, Report 2024-N-6 (Dec. 2025), https://www.osc.ny.gov/files/state-agencies/audits/pdf/sga-2026-24n6.pdf.
EEOC v. Abercrombie & Fitch Stores, Inc., 731 F.3d 1106 (10th Cir. 2013), rev’d, 575 U.S. 768 (2015). Opinions: https://law.justia.com/cases/federal/appellate-courts/ca10/11-5110/11-5110-2013-10-01.html and https://www.law.cornell.edu/supremecourt/text/14-86.
Gonzalez v. Abercrombie & Fitch Stores, Inc., No. 03-cv-02817 (N.D. Cal.). Public case record and documents: https://clearinghouse.net/case/9472/; EEOC, Landmark Resolution (Nov. 18, 2004), https://www.eeoc.gov/newsroom/eeoc-agrees-landmark-resolution-discrimination-case-against-abercrombie-fitch-0.
LinkedIn, How LinkedIn Uses AI Agents to Connect Job Seekers and Hirers, https://www.linkedin.com/help/linkedin/answer/a7437598.
Disney Careers, Candidate AI Usage, https://www.disneycareers.com/en/ai-usage.
42 U.S.C. § 2000e-2(k), (l); EEOC, Section 15 Race and Color Discrimination, part V, https://www.eeoc.gov/laws/guidance/section-15-race-and-color-discrimination.
N.Y. Exec. Law § 296; N.Y.C. Admin. Code § 8-107; NYC, DCWP TLC and CCHR Issue Enforcement Policy Statement on Artificial Intelligence (Oct. 1, 2026), https://www.nyc.gov/site/dca/news/067-26/dcwp-tlc-cchr-issue-enforcement-policy-on-artificial-intelligence. See also DCWP, Automated Employment Decision Tools Frequently Asked Questions, part VII, https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf.
California Civil Rights Department, Civil Rights Council Secures Approval for Regulations to Protect Against Employment Discrimination Related to Artificial Intelligence (June 30, 2025), https://calcivilrights.ca.gov/2025/06/30/civil-rights-council-secures-approval-for-regulations-to-protect-against-employment-discrimination-related-to-artificial-intelligence/.
Legal disclaimer
This article expresses the author’s opinions and provides general legal information as of October 28, 2026. It is not legal advice and does not create an attorney-client relationship. It does not establish that a named company has discriminated against a particular applicant. Legal requirements, deadlines, and remedies depend on the facts, jurisdiction, and current law. Consult qualified counsel regarding a specific matter.
ChatGPT and generative AI disclaimer
This article was developed with assistance from ChatGPT following a discussion with the author. Generative AI can produce inaccurate, incomplete, or outdated information, including mistakes in legal descriptions and citations. The author’s opinions are identified as such; AI assistance is not independent verification or a substitute for professional judgment. Readers should consult the linked primary authorities and original research before relying on a legal or factual proposition.