I think this is expected if you consider how ChatGPT works. It is a language model, so its job is to generate plausible-looking text. It does not use any exter

I think this is expected if you consider how ChatGPT works. It is a language model, so its job is to generate plausible-looking text. It does not use any external reference material during inference so it has to encode all knowledge it uses in its weights. So the weights are essentially its only long-term memory. You could say that the task it was given here was “Provide plausible-looking links without using any reference material.” For that task I think it did a reasonable job.