The sources are there in the training dataset, they are just not linked to the response. I don't think this is an inherent property of LLMs though, and I imagi

The sources are there in the training dataset, they are just not linked to the response. I don't think this is an inherent property of LLMs though, and I imagine future iterations will have some sort of attention mechanism that highlights the contributing source materials.