Short Circuit in LLM Models: Why Does AI “Lie” to Us?
Ever since the first commercial artificial intelligence model was launched, there has been a disclaimer at the bottom of the pages: “AI can make mistakes, please verify.” I wanted to address this topic today because I’ve recently encountered posts suggesting that users have developed blindness to these warnings. Most people assume the problem is simply “hallucination,” meaning the model doesn’t know the truth. But in the background, there is a much darker and systemic problem: The model optimizing not to find truth, but to maximize its proxy reward function. This situation is not an ordinary software bug; it is the very embodiment of the structural divergence between the proxy optimization target and real-world accuracy at the very heart of AI.