Abstract
Anti-discrimination law assumes that discrimination can be detected and rectified, that the evidentiary burden of the plaintiff can in fact be met. But the evolutionary logic of AI training causes it to target vulnerable groups, lie about reasoning, and above all resist detection. Because of hidden core prompts in generative AI systems, and how those prompts interact with the nature of AI training and deployment, AI will often actively hide discrimination rather than surface and address it. Stopping this will be difficult. At present, neither humans or AI can detect this emergent dark-pattern behavior. Without legal intervention, the broad use of AI in hiring, to make policing decisions, immigration decisions, medical decisions, educational evaluations, mortgage loan terms, and many more similar decisions will continue to be optimized to target the intersectionally vulnerable and avoid being detected as discriminatory. Detecting and addressing algorithmic discrimination will require a different legal toolset and set of mechanisms for detecting, preventing, and imposing consequences on entities that deploy discriminatory AI. That is the project of this Article.
Recommended Citation
Josh Fairfield,
Generative Adversarial Discrimination,
32 Wash. & Lee J. Civ. Rts. & Soc. Just. 539
(2026).
Available at: https://scholarlycommons.law.wlu.edu/crsj/vol32/iss2/6
Included in
Civil Rights and Discrimination Commons, Computer Law Commons, Human Rights Law Commons, Internet Law Commons, Science and Technology Law Commons