Access Controls Will Solve the Dual-Use Dilemma

The "dual-use dilemma" creates impossible choices for AI safety systems: Current AI systems cannot distinguish between legitimate and harmful uses of the same information (e.g., chemistry knowledge for research vs. weapons), forcing them to either refuse all requests or allow all requests—both outcomes are problematic since they either block legitimate users or enable harmful ones.

Access controls could replace one-size-fits-all content moderation: Rather than blanket refusals, the paper proposes a framework where verified users (researchers, professionals, etc.) could access sensitive information while unverified users cannot, potentially reducing both "over-refusals" of legitimate queries and "under-refusals" that allow harmful uses.

Current AI systems lack real-world context to make informed decisions: The paper identifies that AI models cannot access information about who is asking, their credentials, or their actual intentions, making it impossible for them to make nuanced judgments about whether a dual-use request is safe or not.

This approach could enable regulators to implement targeted policies: Rather than relying solely on AI companies' internal safety rules, access control frameworks could allow governments and regulators to set specific requirements for who can access certain capabilities, creating a more transparent and accountable system.

This remains a high-level proposal requiring significant implementation work: The authors acknowledge their framework is conceptual and would require developing verification systems, institutional partnerships, and technical infrastructure before it could be practically deployed.