时文摘要
我们能否全然信任人工智能输出的结果？剑桥大学等高校研究表明，面对部分复杂系统，即便拥有海量数据，AI也难以可靠求解，还会出现偏移、生成虚假信息等问题，科研团队还研发出新型算法应对该难题。
Can we always trust the results we get from AI? Researchers from the University of Cambridge and the University of California Santa Barbara say no. They have shown that there are some problems that even the most powerful AI cannot reliably solve, no matter how much data it is given.
Many real-world systems, like those in oceans, the human brain, or robots, are too complex to describe neatly with equations (方程式). So researchers often use machine learning to study how they behave. But these AI methods do not always work well. Sometimes they return unreliable results or poor predictions. Sometimes, however, providing reliable solutions may be fundamentally impossible for some problems, even with infinite data.
The researchers designed special systems to test AI. These systems were built to find out exactly where and why AI prediction breaks down. They identified two main reasons why machine learning fails on complex systems. Either the algorithm cannot tell when it has seen enough data, or patterns in the system are hidden and hard to distinguish.
Their results may also help explain why AI chatbots can be accurate in the short term but drift or produce false information over time. The researchers found that chaotic systems are especially problematic. When a system is chaotic, meaning tiny differences in starting conditions lead to very different results, short-term prediction can be accurate, but long-term prediction becomes unreliable. Small changes in a question can send a chatbot down a completely different path. The answer looks reasonable word by word but produces unrealistic information over longer outputs.
