Even with that court ruling, I would suspect that what Frederick intends to build could be legal for the same reason that search engines don't innately infringe on copyright. Fair use allows you to list and cite works and to provide a brief summary, which is why Google is able to provide you with a link to a website without getting sued, even if they don't have the rights to any of the site's content.

It would be much more questionable to build more extensive "AI summaries" of the C++ papers though. Anyhow, I'm not a legal expert, and this is all a bit of a gray area still until more court rulings.

Even if it was perfectly legal, it's morally questionable to train an AI search engine for C++ papers when a substantial portion of authors would take issue with it.

However, I find the idea to be generally useful, and maybe there's a version of this which authors would find much less objectionable. For example, it would be nice if you could search papers by the standard library symbols or keywords that they've added; e.g. if you could find P3411R2 by searching for "std::any_view", or P1024R3 by searching for "std::span::front". Paper tagging seems much more obviously like fair use than turning papers into an AI chatbot, but it's still a good use of machine learning.