Weak AI regulation could make artificial intelligence products less safe than having no regulation at all, according to a new study by researchers from Cornell University and Carnegie Mellon University.
The research, published in the Proceedings of the National Academy of Sciences, examined how different regulatory approaches affect companies involved in developing and deploying AI systems. The researchers focused on two groups: companies that build general-purpose AI models and downstream firms that adapt those models for specific applications.
Researchers found that the existence of safety regulations, applicable to the downstream firms, could cause an unintentional ‘free-riding’ effect. There can be little motivation for general AI creators to invest in safety measures, such as conducting external audits. They could instead rely on downstream companies to address safety risks when adapting the technology for specific uses.
The research model defined safety broadly. Risks included potential harm to users, such as toxic or harmful responses from AI-powered systems. The researchers said the findings highlight the importance of examining incentives across the entire AI supply chain.
The study also identified a potential ‘sweet spot’ in regulation. Both general-purpose AI developers and downstream companies need defined safety standards, which could improve outcomes for consumers and businesses.
According to the researchers, Shared safety requirements reduce uncertainty between companies. Each party can make safety investments with greater confidence that the other side will also meet its obligations.
The researchers said this approach could improve both safety and performance while potentially increasing profits for the companies involved. The study comes as governments continue to debate how to regulate rapidly evolving AI systems. Looking ahead, experts warned against treating AI as a single product controlled by one company.
Also Read: Alphabet Develops ‘Frozen v2’ to Make Gemini Faster