But I think the worst part of this is that certain kids still need individual attention even if they aren't special needs and there is no way the AI will be able to pick up on that or act on it.
Teachers already miss special needs students all the time. If anything, an AI's pattern recognition will likely be more able to detect areas a student struggles in, because it can analyze a student's individual performance in a sandbox. Teachers have dozens of students to keep track of at any given time, and it's impossible for them to catch everything because we feeble humans have limited mental/emotional bandwidth, unlike our perfect silicon gods.
The truth is that this will actually do a lot of things better than real teachers. It'll also do a lot of things worse. It'll be interesting to see how the trade-off plays out and to see which elements of the project are successful enough to incorporate into traditional learning environments.