A theoretical modeling study from Cornell University and Carnegie Mellon University reveals that poorly designed artificial intelligence regulation can paradoxically increase safety risks, potentially causing more harm than having no oversight at all. Published in the Proceedings of the National Academy of Sciences, the research addresses growing concerns as fragmented state-level laws attempt to fill the vacuum left by absent federal AI standards. Led by doctoral researcher Benjamin Laufer alongside Cornell Professor Jon Kleinberg and Carnegie Mellon Assistant Professor Hoda Heidari, the study introduces a computational framework to map regulatory incentives across the AI development pipeline. The model differentiates between upstream entities that train general-purpose foundation models and downstream firms that integrate those models into specialized applications like customer service platforms or clinical diagnostics. By simulating varying safety thresholds for each tier, the researchers quantified how regulatory design directly influences corporate investment in safety audits and risk mitigation. The analysis uncovered a counterintuitive backfiring effect. When legislation imposes minimal safety obligations exclusively on downstream developers, foundational model providers face reduced incentives to implement rigorous safety protocols. Instead, upstream companies strategically offload compliance burdens, resulting in a net decrease in product safety compared to an unregulated market. Conversely, the model identifies a regulatory sweet spot where both upstream and downstream entities are held to coordinated safety targets. This dual-tier requirement fosters mutual accountability, lowers systemic risk, and ultimately generates higher commercial returns alongside improved consumer protection. The findings underscore the necessity of supply-chain-wide regulatory design. Policymakers aiming to establish federal AI standards must account for the complex, multi-stakeholder nature of model development rather than targeting isolated endpoints. While the current framework operates as a simplified theoretical construct, the research team plans to validate its predictions against real-world regulatory implementations and expand the simulation to accommodate global jurisdictions with divergent compliance standards. The study ultimately provides a data-driven foundation for crafting legislation that aligns corporate incentives with public safety, ensuring that AI oversight strengthens rather than undermines industry innovation.