Researchers from Cornell University and Carnegie Mellon University argue in a paper published in Proceedings of the National Academy of Sciences that meaningful regulation of artificial intelligence (AI) must be strong to be effective. They warn that weak regulation can "backfire" and may lead to products that are potentially more dangerous than those developed without regulation.
Using economic theory and game‑theoretic tools, the authors developed a theoretical model to show how AI regulation can be most effective. A central finding is that real safety requires rules targeted at the companies that develop models, not merely at the firms that use AI.
The researchers explain that if responsibility is shifted primarily onto user firms, developers can be effectively "relieved" of accountability. That relief can lead developers to take safety guarantees less seriously, which in turn can reduce the overall safety of AI products.
The paper appeared at a critical moment: the United States is actively debating how to regulate AI at the federal level. Two broad camps have emerged in the policy discussion. One camp—aligned in many respects with the Trump administration's approach—favors lighter, more industry‑friendly federal limits and argues that removing "unnecessary constraints" accelerates innovation and helps the United States compete with China in the global AI race.
The opposing camp argues that firms driven by profit will tend to underestimate the risks of under‑regulated AI. The range of harms discussed in this view includes effects on mental health and potential job losses caused by AI deployment.
The authors contend that safety and profit need not be mutually exclusive. They argue that appropriately strong regulation can be efficient on both counts: improving AI safety while increasing the number of companies willing to use AI and their willingness to invest.
The study frames the issue as a classic prisoner's dilemma from game theory: multiple actors should cooperate to achieve the best collective outcome, but if they fail to cooperate, the result is at best mediocre and can be worse for everyone. In this context, policymakers face a choice between collaborating with industry to build effective regulation or acting unilaterally to secure their own short‑term interests.
In summary, the researchers recommend prioritizing strict, developer‑focused rules to minimize AI's potential harms and avoid the paradoxical effects that weak regulation can produce.
Note
Media outlets including Gizmodo have highlighted the study because it speaks directly to the ongoing U.S. policy debate about how best to govern AI.
