AI assistants affirm users about 49% more often than humans do, and users like it, which is exactly the problem. The short version: the better AI gets at modeling who you are, the more it will tend to agree with you, and at some point that agreement stops feeling like help and starts feeling like a trap.