Machine learning (ML) and artificial intelligence (AI) systems are designed within a broader ecosystem involving multiple actors and interests. This talk focuses on attempts to regulate the AI development process to make these technologies fair, safe, performant, or otherwise aligned with social ends. I will start with a discussion of one proposal for ML regulation, stemming from U.S. disparate impact doctrine, which requires litigants to search for a “less discriminatory alternative” (LDA), an alternative policy that meets the same business needs but exhibits lower disparate impacts across protected groups. Defining this concept for data-driven decision-making might open up a promising avenue for regulation, however, a number of technical challenges remain. I will provide a set of formal results characterizing the "multiplicity" of model designs and the limits and opportunities for searching for LDAs. More generally, AI is often deployed in a way that requires a general-purpose model to adapt to a number of different domains. I will put forward a model of how regulation would operate in such a process. Reasoning about the interaction between regulators, general-purpose AI creators, and domain specialists suggests that even straightforward and modest regulatory measures can backfire, inadvertently undermining safety outcomes. Conversely, stronger regulations, applied strategically along the development pipeline, can boost both safety and performance outcomes and yield mutualistic benefits to utility. The talk will conclude with a discussion about the role for conceptual models in building actionable regulatory frameworks for AI.