But you can enter in real search data as a prompt, and use its training to summarize it. (Or it can fill its own prompt automatically from an automatic search)

It won't/can't update it's priors, and I agree with you there, but it can produce novel output on a novel prompt with its existing model/weights