For some questions that refer to old knowledge, such as questions that have been studied for a very long time, the AI still looks up recent papers (up to 2000), which is likely because previous versions are not widely available in a digital text format, but instead in a PDF.
Perhaps in the cases when it is clear that it is a long-studied topic, the AI should choose papers based not on their top position, which usually indicates the novelty they bring, but instead also based on the number of references or another heuristic, that allows identifying papers with a good overview of the field, such as literature review papers. If not done, the AI struggles to understand the topic well as it does not read the foundational papers, jumping straight to the cutting edge.
For some questions that refer to old knowledge, such as questions that have been studied for a very long time, the AI still looks up recent papers (up to 2000), which is likely because previous versions are not widely available in a digital text format, but instead in a PDF.
Perhaps in the cases when it is clear that it is a long-studied topic, the AI should choose papers based not on their top position, which usually indicates the novelty they bring, but instead also based on the number of references or another heuristic, that allows identifying papers with a good overview of the field, such as literature review papers. If not done, the AI struggles to understand the topic well as it does not read the foundational papers, jumping straight to the cutting edge.