I feel like there's too much going on for people unfamiliar with the topic. The "people in a box" example sort of comes out of nowhere, and may cause your audience to think anthropomorphically. I think you should just evade the issue of how much smarter than us your hypothetical AI is, and focus on getting the audience to understand that there are seriously-considered AI decision-making systems that lead to very bad things. Using examples is good, but I wish you allotted more time to explain yours - as it was, the conclusions just seemed asserted. Were I to give a talk like this, I would actually cut down to one general example (probably reinforcement learning, which is useful for dumb robots but a super bad idea for smart robots) and go into it in depth, rather than using several fun examples, but I think that's mostly a stylistic preference. Also, kudos for getting the audience to laugh :)