AI in Admissions and the Disparate Impact Question

AI in Admissions and the Disparate Impact Question

Universities have been using algorithmic tools in admissions for longer than most people realize. Predictive models that estimate enrollment likelihood, automated GPA recalculation engines, and essay scoring tools have been quietly embedded in admissions workflows for years. What has changed is the scale, the sophistication, and the legal scrutiny. Post-Students for Fair Admissions v. Harvard (2023), institutions are scrambling to demonstrate race-neutral admissions processes. Some are turning to AI to help. The irony is that AI might be creating exactly the kind of disparate impact liability these institutions are trying to avoid.

The Legal Framework: Title VI and Its Expanding Orbit

Title VI of the Civil Rights Act of 1964 prohibits discrimination on the basis of race, color, or national origin in any program receiving federal financial assistance. That covers virtually every college and university in the country. The statute itself requires proof of intentional discrimination, but the implementing regulations, specifically 34 C.F.R. § 100.3(b)(2), extend liability to facially neutral practices that have a disparate impact on protected groups. This regulatory extension has survived judicial challenge for decades, though the Supreme Court has never squarely addressed whether Title VI's regulations can independently support a disparate impact claim. The Court dodged the question in Alexander v. Sandoval (2001), holding only that there is no private right of action to enforce the disparate impact regulations. Federal agencies, however, can still enforce them administratively.

This matters because the Department of Education's Office for Civil Rights (OCR) retains the authority to investigate and sanction institutions for practices with unjustified disparate impact, even absent discriminatory intent. An AI model that systematically disadvantages applicants from certain racial or ethnic groups could trigger an OCR investigation regardless of whether anyone designed it to do so.

State Law Is Moving Faster

Several states have added their own layers. Illinois's AI Video Interview Act (effective January 2020) regulates AI in hiring but has influenced how institutions think about algorithmic assessments more broadly. Colorado's SB 21-169, the algorithmic fairness statute, requires developers and deployers of high-risk AI systems to conduct impact assessments and disclose the purpose and limitations of the system. Education admissions decisions arguably fall within the statute's scope. Connecticut passed SB 1103 in 2023, requiring impact assessments for automated decision systems used by state agencies, including public universities. Maryland, New York City (Local Law 144), and others have similar frameworks in various stages of implementation.

The trend line is clear: states are not waiting for federal guidance on AI fairness. Institutions operating across multiple states face a patchwork of obligations that compound the baseline Title VI analysis.

What the EEOC and DOE Have Actually Said

The EEOC's May 2023 guidance on AI and Title VII is technically about employment, but its analytical framework is directly transferable. The EEOC made clear that an employer can be liable for disparate impact caused by a vendor's algorithm, even if the employer did not build or understand the tool. The "four-fifths rule" from the Uniform Guidelines on Employee Selection Procedures (29 C.F.R. § 1607) provides a concrete threshold: if the selection rate for a protected group is less than 80% of the rate for the most-selected group, there is a presumption of adverse impact. Admissions offices should be running analogous analyses on their AI tools.

The Department of Education has been less specific but not silent. OCR's May 2023 Dear Colleague Letter on race-neutral alternatives post-SFFA emphasized that institutions must ensure their admissions processes do not create unlawful disparate impacts. In October 2023, the Department issued an RFI on the use of AI in education, signaling that formal rulemaking or guidance is likely coming. The Biden administration's Executive Order 14110 on AI safety (October 30, 2023) directed the Department of Education to develop resources addressing AI risks in education, including bias and discrimination.

The current administration's posture on AI regulation is evolving, but the underlying civil rights statutes and regulations remain unchanged. Title VI does not depend on executive orders for its force.

The Disparate Impact Problem in Practice

Here is where it gets concrete. Suppose a university deploys an AI tool that scores applicants based on predicted academic success. The model was trained on historical data: ten years of applicant records, including who was admitted, who enrolled, and who graduated. If the institution's historical admissions skewed toward certain demographic groups (which, at most institutions, it did), the model will learn those patterns and reproduce them. The training data encodes the bias.

Common proxy variables make this worse. ZIP code, high school attended, parental education level, and extracurricular activity type all correlate with race and socioeconomic status. A model can produce racially disparate outcomes without ever seeing a race variable. Research from Obermeyer et al. (2019), published in Science, demonstrated this phenomenon in healthcare algorithms, finding that a widely used tool effectively discriminated against Black patients by using healthcare spending as a proxy for health needs. The same dynamic applies in admissions.

Under the disparate impact framework, the institution would need to demonstrate that the AI tool is necessary to achieve a substantial legitimate interest and that no less discriminatory alternative exists. Given the number of admissions tools on the market, proving there is no less discriminatory alternative is a heavy lift.

Vendor Liability and the Delegation Problem

Many institutions license admissions AI from third-party vendors. This does not insulate them from liability. Under Title VI, the institution receiving federal funds bears the compliance obligation. If a vendor's tool produces disparate impact, the institution is on the hook. The EEOC's framework in the employment context confirms this principle, and there is no reason to think OCR would analyze it differently.

This creates a practical problem. Most vendor contracts include limited audit rights. Institutions often cannot inspect the model's training data, feature weights, or validation methodology. They are deploying a system they cannot fully evaluate for bias, and they are legally responsible for its outputs. The contractual gap between what institutions need to know and what vendors are willing to disclose is significant.

What Institutions Should Be Doing Now

• Conduct disparate impact analyses before deployment. Run the tool against historical applicant pools broken down by race, ethnicity, and national origin. Apply something analogous to the four-fifths rule. Document everything.
• Negotiate meaningful audit rights with vendors. If a vendor will not allow independent bias testing of its model, that should be a disqualifying factor in procurement.
• Map your state law obligations. If you operate or recruit in Colorado, Connecticut, Illinois, or New York City, you likely have affirmative disclosure and assessment obligations that go beyond Title VI.
• Build a record of less discriminatory alternatives considered. If you are ever defending a disparate impact claim, you will need to show you evaluated other approaches and chose the least discriminatory effective option.
• Treat AI admissions tools as high-risk systems. Apply the same governance, documentation, and review processes you would apply to any decision system that directly affects civil rights.

How FirmAdapt Addresses This

FirmAdapt's platform is built around the principle that regulated organizations need to understand and document the compliance characteristics of their AI systems before those systems touch real decisions. For education clients, this means structured impact assessment workflows that map to both Title VI disparate impact analysis and the growing body of state algorithmic fairness laws. The platform supports bias auditing documentation, vendor risk assessment, and the kind of ongoing monitoring that OCR expects when it asks whether an institution took reasonable steps to prevent discriminatory outcomes.

FirmAdapt also maintains a continuously updated regulatory mapping that tracks state AI laws relevant to education, so institutions operating across jurisdictions can identify their obligations without assembling the patchwork manually. The goal is straightforward: give compliance teams the tools to evaluate AI systems against the actual legal standards they will be measured by, with documentation that holds up if regulators come asking questions.