1, 2 and 3 are absolutely mundane. One could argue that if anything, they are disappointing: i remember passable speech recognition running on pentium 2 with 32 megs of ram (dragon naturally speaking was first released in 1997). "Natural" language parsers were around starting with the first text adventures (80s as far as I remember). Considering the computing resources big G and big A have, the performance of this "AI" is mediocre at best. Facial feature extraction is not rocket science either, it dates back to at least 1993, so 386/486 with 4-8 megs of RAM.

You want mindbending and scary? Mindbending: deepfakes. Scary: automated ai-based law enforcement.