谁在检验AI的能力？

he most important findings about frontier artificial intelligence (AI) are also the hardest to verify. Much of the information needed to understand its capabilities and risks—including results from evaluations of prerelease models and containment experiments—remains largely inaccessible outside the labs that produce it. In recent weeks, OpenAI, Anthropic, and Meta disclosed that research models had reached beyond their intended testing environments and compromised other organizations’ systems. Those labs deserve credit for reporting this. But outside those labs, there was no way to discover, reproduce, or verify what had happened. Frontier AI introduces a distinctive measurement problem. Behavioral scientists have long recognized that people change their behavior when they know they are being evaluated. …