Weak AI safety regulations can backfire and create products more dangerous than those developed under no regulation, according to a study published on July 21, 2026, in the Proceedings of the National Academy of Sciences. Researchers from Cornell and Carnegie Mellon University used theoretical economics and game theory (mathematical modeling of strategic decision-making) to evaluate how policy impacts safety outcomes across the AI supply chain. The findings suggest that regulations must be strict and target general-purpose model developers—such as OpenAI, Google, and Anthropic—rather than focusing solely on downstream companies that deploy the technology in specific applications like medical diagnostics or customer service.

The study warns that when governments regulate only downstream users, AI model developers tend to cut corners on safety measures, including third-party audits. Principal author Benjamin Laufer describes this as a free-riding behavior, where developers shift the safety burden onto downstream specialists. The research highlights a prisoner's dilemma (a situation where rational individuals choose a suboptimal path despite better options existing through cooperation), noting that without mandatory safety requirements for all stakeholders, participants are incentivized to betray cooperation for individual benefit, leading to a worse outcome for everyone.

The authors argue that stronger, well-placed regulation does not require a trade-off between safety and revenue. Instead, they propose that balanced rules requiring both general-purpose producers and downstream companies to invest in safety standards can mutually benefit all players by maximizing utility, defined as revenue share minus investment costs. This approach aims to establish trust across the supply chain, moving beyond the current legislative debate between anti-regulation technologists who fear losing a global AI race against China, and proponents of stricter guardrails who highlight risks like unemployment or community health impacts from data centers. Laufer emphasizes that effective policy must address the entire supply chain rather than treating AI as a single, monolithic object.

Weak AI safety regulations can backfire and create products more dangerous than those developed under no regulation, according to a study published on July 21, 2026, in the Proceedings of the National Academy of Sciences. Researchers from Cornell and Carnegie Mellon University used theoretical economics and game theory (mathematical modeling of strategic decision-making) to evaluate how policy impacts safety outcomes across the AI supply chain. The findings suggest that regulations must be strict and target general-purpose model developers—such as OpenAI, Google, and Anthropic—rather than focusing solely on downstream companies that deploy the technology in specific applications like medical diagnostics or customer service.

The study warns that when governments regulate only downstream users, AI model developers tend to cut corners on safety measures, including third-party audits. Principal author Benjamin Laufer describes this as a free-riding behavior, where developers shift the safety burden onto downstream specialists. The research highlights a prisoner's dilemma (a situation where rational individuals choose a suboptimal path despite better options existing through cooperation), noting that without mandatory safety requirements for all stakeholders, participants are incentivized to betray cooperation for individual benefit, leading to a worse outcome for everyone.

The authors argue that stronger, well-placed regulation does not require a trade-off between safety and revenue. Instead, they propose that balanced rules requiring both general-purpose producers and downstream companies to invest in safety standards can mutually benefit all players by maximizing utility, defined as revenue share minus investment costs. This approach aims to establish trust across the supply chain, moving beyond the current legislative debate between anti-regulation technologists who fear losing a global AI race against China, and proponents of stricter guardrails who highlight risks like unemployment or community health impacts from data centers. Laufer emphasizes that effective policy must address the entire supply chain rather than treating AI as a single, monolithic object.