I see AI as breaking the traditional contract between humans and computers. It used to be that, if a computer was wrong, it was our fault. Maybe bad software. Maybe bad hardware. Maybe bad shielding from gamma rays, but it was our fault. We've now started using computers against ultra-complex problems that often don't even have a true "right" solution, or have chaotic enough input that it's impossible to be "smart enough" to always be right. Essentially, we've made computer be as fallible as us, and I don't think we understand the consequences of that yet.

I think we're going to find that there's a soft limit, or at least significant diminishing returns, on how reliable AI can be. There will be things it will be better at (it excels at having a broad knowledge base), but it can't understand what makes a good human experience better than a human can (duh!).

Coding seems like an ideal application for AI. It's wicked complex, but also generally provably right or wrong to some degree (i.e. does it work?). It won't be solving world peace any time soon because A) there is no solution, and B) even getting closer to a solution requires a deep understanding of the human factors.