Nice, Peter, thanks for considering it! You see, the prompt can be handled to not change the author’s tone and maintain the result only from the specified data from the links.
That is the approach used at https://ask.discourse.com, where the responses are tightly bound to docs and a curated base. The model does not use its training data almost at all, and I think that is the most clever and productive way to actually use the technology.
There also are inexpensive models that are really good at summarizing. So I hope it will be implemented; it can improve the reader experience and also the SEO/IA-SEO practices as well.