AI Bias and Discrimination Law: Legal Risks for Businesses Using AI

One of the most significant legal risks associated with deploying artificial intelligence for consequential decisions is the risk of discrimination. AI systems are trained on historical data, and historical data reflects historical patterns of human decision-making — patterns that have often been shaped by discrimination, inequality, and systemic bias. When an AI learns from this history and then applies what it has learned to new decisions, it can replicate and in some cases amplify the discriminatory outcomes embedded in its training data. For businesses, this is not merely an ethical concern. It is a legal liability that existing federal and state anti-discrimination laws address directly.

Disparate Impact: The Core Legal Theory

The primary legal theory under which AI discrimination claims arise is disparate impact, also called adverse impact. Disparate impact occurs when a facially neutral policy or practice — one that does not explicitly discriminate — disproportionately harms members of a protected group. Unlike disparate treatment claims, which require proof of intentional discrimination, disparate impact claims focus on the outcomes of a practice, not the intent behind it.

Courts and regulatory agencies have consistently held that disparate impact is a cognizable legal theory under Title VII of the Civil Rights Act (employment discrimination), the Fair Housing Act (housing discrimination), the Equal Credit Opportunity Act (credit discrimination), and several other federal laws. The Supreme Court affirmed the disparate impact theory’s role in fair housing law in Texas Department of Housing and Community Affairs v. Inclusive Communities Project in 2015. The existence of this theory means that a business can face discrimination liability even if its AI system makes decisions without any explicit consideration of race, gender, age, or other protected characteristics — as long as the system’s decisions have a disproportionate adverse impact on members of a protected group.

AI in Employment: EEOC Guidance and the Four-Fifths Rule

The Equal Employment Opportunity Commission has made clear through technical assistance documents and public guidance that employers are responsible for the discriminatory effects of AI tools they use in employment decisions, regardless of whether the employer developed the tool itself. If an employer uses an AI-powered applicant screening tool, resume parser, video interview analyzer, or any other algorithmic system that has a disparate impact on a protected group, the employer — not just the vendor — faces potential liability under Title VII, the Age Discrimination in Employment Act, and the Americans with Disabilities Act.

The EEOC’s longstanding Uniform Guidelines on Employee Selection Procedures, which predate AI but apply fully to algorithmic selection tools, establish the four-fifths rule as a practical test for adverse impact. Under the four-fifths rule, if the selection rate for a protected group is less than four-fifths (80%) of the selection rate for the group with the highest selection rate, adverse impact is indicated. For example, if 50% of white male applicants pass an AI screening tool but only 30% of Black female applicants pass, the Black female selection rate (30%) is less than 80% of the white male rate (40%), indicating adverse impact. Employers should apply this kind of analysis to any AI tool used in hiring, promotion, or other employment decisions.

Beyond the four-fifths rule, the EEOC has indicated that AI tools that disadvantage job applicants with disabilities may violate the ADA if they fail to reasonably accommodate individuals who need alternative testing formats or who are disadvantaged by the technology itself. AI video interview analysis tools, for example, may penalize candidates who have certain speech disorders, visual impairments that affect eye contact, or neurodivergent characteristics that present differently in video interviews.

AI in Lending: Fair Lending Law and Model Risk

Credit and lending decisions made using AI face oversight under the Equal Credit Opportunity Act and the Fair Housing Act. The CFPB has issued guidance emphasizing that lenders must be able to explain specific reasons for adverse credit actions to applicants even when an AI model made the decision. A lender cannot satisfy the adverse action notice requirement of ECOA by citing the AI’s algorithm as the reason — the notice must include specific, principal reasons tied to the applicant’s actual characteristics and credit history.

The CFPB and federal banking regulators have also emphasized model risk management for AI-based credit models. Lenders are expected to conduct regular validation of their credit models to identify whether they produce discriminatory outcomes across demographic groups. The use of machine learning models in credit scoring is permissible, but those models must be subject to the same fair lending standards that apply to traditional credit scoring, and lenders must have the analytical capability to identify and address disparate impact when it arises.

The use of alternative data in AI credit models — social media activity, online purchasing behavior, device usage patterns, and other non-traditional data points — is a particular area of regulatory scrutiny. Alternative data may appear neutral but can correlate strongly with race, national origin, or other protected characteristics in ways that create discriminatory credit outcomes. Lenders incorporating alternative data into AI models must carefully evaluate the fair lending implications of those data elements.

AI in Housing: Fair Housing Act Concerns

AI tools used in rental and sales housing decisions — applicant screening, pricing algorithms, chatbot interactions with prospective tenants or buyers — are subject to the Fair Housing Act, which prohibits discrimination in housing on the basis of race, color, national origin, religion, sex, familial status, and disability. The Department of Housing and Urban Development has issued guidance indicating that algorithmic tenant screening tools may violate the Fair Housing Act if they disproportionately screen out applicants of protected races or national origins based on criteria that are not justified by legitimate business necessity.

Advertising and recommendation algorithms also raise fair housing concerns. Housing platforms that use AI to target advertising or recommend listings in ways that effectively steer members of protected groups toward or away from certain neighborhoods may be engaged in digital redlining. The National Fair Housing Alliance and other organizations have documented algorithmic discrimination in housing advertising, and regulators have brought enforcement actions against platforms that use targeting criteria that function as proxies for protected class membership.

State AI Non-Discrimination Laws

In addition to federal anti-discrimination laws, several states have enacted specific requirements relating to algorithmic discrimination. Colorado’s AI Act, which took effect in 2026, explicitly requires developers and deployers of high-risk AI systems to use reasonable care to avoid algorithmic discrimination based on protected characteristics and provides for enforcement by the Colorado Attorney General. Illinois has amended its Human Rights Act to address employment AI discrimination specifically. New York City enacted Local Law 144, which requires employers to audit algorithmic employment tools for bias before using them in hiring or promotion decisions and to provide notice to applicants that such tools are being used.

Practical Steps for Businesses

Businesses using AI for consequential decisions in employment, lending, housing, healthcare, or other regulated contexts should take several practical steps to assess and manage their bias risk. Before deploying any AI system that affects individuals in a regulated context, request from the vendor demographic performance data showing the system’s selection or approval rates across racial, gender, and age groups. If the vendor cannot provide this data, that itself is a significant red flag. Conduct your own adverse impact analysis using the four-fifths rule or other appropriate statistical methodology after deployment and periodically thereafter.

Ensure that AI vendor agreements include representations about the system’s compliance with applicable anti-discrimination laws and, ideally, indemnification for discrimination claims arising from the system’s design or training. Establish human review processes for consequential AI-driven decisions, particularly adverse decisions, so that a human being can assess whether any AI recommendation appears inconsistent with applicable law. Document your evaluation, monitoring, and review processes so that you can demonstrate reasonable care and due diligence if a discrimination claim arises.

The legal risk from AI-driven discrimination is real, the regulatory attention is intensifying, and the consequences of getting it wrong — EEOC charges, class actions, regulatory enforcement, reputational damage — can be severe. Businesses that invest in bias evaluation and monitoring before they experience a problem are in a substantially better position than those that wait for an adverse event to prompt action.