AI Contamination

AI applications have proven their superiority in specific tasks, such as chess and Go, over human capabilities. Like ChatGPT, more general AI systems have also made significant strides, demonstrating impressive capabilities in language translation, writing assistance, answering questions, and providing educational support, among other areas. As we embrace these advancements, it’s essential to be mindful of the potential issues they may introduce. One of such problems is AI contamination. There are two main issues with AI training and applications. Large AI models, such as ChatGPT, excel at handling common tasks based on well-established knowledge. However, when faced with more advanced topics, especially those at the forefront of research where knowledge is unavailable or not well established, these models tend to generate content that is not scientifically verified or proven, rather than acknowledging the lack of reliable information. A further concern is that individuals, such as particularly students, who lack reliable background knowledge may accept AI-generated content as the absolute truth. This can mislead them directly and shape the content they produce in the future. When subsequent AI models are trained on such flawed material, these misconceptions are reinforced and propagated more widely. As a result, the collective knowledge base becomes contaminated, and the reliability of future models may be progressively undermined.