人工智能公司遗忘的智能：为何AI无法说“我不知道”

The current landscape of artificial intelligence, heavily driven by the pursuit of benchmark scores and performance metrics, has inadvertently overlooked a crucial aspect of intelligence: the ability for AI systems to admit "I don't know." This deficiency stems from prevailing model architectures and training methodologies that prioritize generating a plausible answer in all scenarios, rather than recognizing and communicating the boundaries of their knowledge. An AI system capable of effectively expressing uncertainty would significantly enhance its trustworthiness and reliability, particularly in high-stakes applications such as healthcare and finance. Developing AI that can perform metacognitive assessments and quantify its own confidence is a critical future direction. This capability would prevent the dissemination of misleading information and foster more efficient and safer human-AI collaboration. Achieving this requires moving beyond mere performance optimization to a deeper form of intelligence construction, empowering AI to understand its own limitations. The focus should shift from simply providing an answer to providing a well-calibrated answer, including an explicit indication of when the system lacks sufficient information or certainty. This evolution is essential for building more robust, transparent, and genuinely intelligent AI systems that can operate effectively and responsibly in complex real-world environments.