Rand兰德：2026年AI反扩散经济学：典型化经济分析框架报告（英文版）

■Controls on dangerous artificial intelligence (AI) capabilities can weaken on their own, even when policy stays unchanged. Once a capability has been demonstrated, the cost of reproducing it tends to fall. Each drop in cost pulls new actors past the point at which the capability is worth more than it costs to acquire.A regime that holds its rules fixed must therefore govern a steadily growing pool, and the members of that pool gain more from evasion as acquisition gets cheaper. Thus, a regim