What does NJ DCR Algorithmic Discrimination Guidance (January 2025) say about AI in banking?

What are the key points of NJ DCR Algorithmic Discrimination Guidance (January 2025)?

• Issued by the Office of the Attorney General and the Division on Civil Rights, dated January 2025 and announced January 9, 2025 as part of the Division's Civil Rights and Technology Initiative, which also launched a Civil Rights Innovation Lab.
• Definition: an 'automated decision-making tool' is any technological tool, including software, a system or a process, used to automate all or part of the human decision-making process, from generative AI and machine-learning models to traditional statistical tools and decision trees.
• Part II: the LAD 'prohibits algorithmic discrimination in employment, housing, places of public accommodation, credit, and contracting' on the basis of actual or perceived race, religion, color, national origin, sexual orientation, pregnancy, breastfeeding, sex, gender identity, gender expression, disability and other protected characteristics.
• Disparate treatment: designing or using a tool to treat a protected class differently, or selectively using a tool on one class, violates the LAD; so can recommendations based on a close proxy for a protected characteristic (the guidance's example is requiring individual taxpayer identification numbers instead of Social Security numbers).
• Disparate impact: a tool that disproportionately harms a protected class is prohibited unless it serves a substantial, legitimate, nondiscriminatory interest, and even then if a less discriminatory alternative exists; whether the entity tested its tool for bias or evaluated alternatives may be considered as relevant evidence.
• Liability: a covered entity is not shielded because a third party developed the tool or because it does not understand the tool's inner workings; the guidance cites the CFPB's Circular 2022-03 on adverse-action notices for credit decisions based on complex algorithms for the point that complexity does not excuse ECOA and Regulation B compliance.
• Reasonable accommodations: algorithmic tools that preclude or impede accommodations for disability, religion, pregnancy or breastfeeding can violate the LAD.
• Footnote 1 states the guidance imposes no new or additional requirements; footnote 51 lists non-exhaustive mitigation steps: quality control of data, impact assessments, pre- and post-deployment bias audits by independent parties, notice of use of a tool, involving affected people in development, and red-teaming.

What did NJ DCR Algorithmic Discrimination Guidance (January 2025) change for banks?

New Jersey had no AI-specific consumer law, and this guidance made clear that the state's civil-rights enforcer will apply the existing LAD to automated tools in credit and other areas on a disparate-impact basis, including for tools bought from vendors. It became more consequential with the Division's disparate-impact rules (N.J.A.C. 13:16), adopted November 5, 2025 and effective December 15, 2025, which codify the burden-shifting standard the guidance describes and cover housing financial assistance and automated employment decision tools. For banks, the change is that a vendor's model is the bank's own exposure under the LAD, and testing for bias is both a defence and a compliance control.

What does New Jersey's algorithmic discrimination guidance require of banks and lenders?

The January 2025 guidance requires nothing new; it explains that the Law Against Discrimination already applies to automated decision-making tools in credit, employment, housing, public accommodation and contracting. For a bank or lender it means the tool-level conduct is judged like any other practice: intentional or proxy-based differential treatment violates the LAD, and a facially neutral model that disproportionately harms a protected class is unlawful unless it is necessary to achieve a substantial, legitimate, nondiscriminatory interest and no less discriminatory alternative would serve the same interest. The guidance treats bias testing and evaluation of alternatives as relevant evidence, and it says liability is not avoided by using a vendor's tool or by not understanding how it works. The Division's disparate-impact rules (N.J.A.C. 13:16), in effect since December 15, 2025, now set the burden-shifting framework.

The guidance sets out why algorithms discriminate in three stages: design (what the model is asked to predict, which inputs it uses), training (skewed, unrepresentative or historically biased data) and deployment (use in ways a tool was not designed for, feedback loops). In the credit context it notes both the risk and the promise, citing consumer-advocate and fintech arguments that non-traditional credit factors can raise approval odds for people with low traditional scores. It does not say any particular model is unlawful; its examples are illustrations, not findings of violations.

The liability discussion matters most for banks buying models. The Division says covered entities often rely largely or entirely on third-party developers, and that this is no shield; it also says developers and vendors of these tools should evaluate design and testing carefully. That parallels the position federal regulators have taken under ECOA and Regulation B, which the guidance cites through the CFPB's 2022 circular on complex algorithms.

Because the guidance is explanatory, enforcement runs through the LAD and, since December 15, 2025, N.J.A.C. 13:16. Banks should treat the guidance as the Division's interpretive view and the rules as the operative standard. Neither document addresses whether federally chartered institutions are subject to the LAD in a given context, which is a preemption question for counsel.

WHAT THIS MEANS IN PRACTICE

• Include disparate-impact testing and a documented search for less discriminatory alternatives in the validation of every credit, pricing and marketing model that reaches New Jersey customers.
• Check vendor contracts for the testing documentation, data descriptions and audit access you would need to defend the model to the Division, since reliance on a vendor is not a defence.
• Screen model inputs and knockout rules for close proxies for protected characteristics such as national origin or source of income.
• Make sure automated application and servicing channels have a route to request a reasonable accommodation.
• Keep records of impact assessments, bias audits and design choices; the guidance treats testing as relevant evidence in a disparate-impact claim.