If you can get a user to think, “I want pizza delivered,” rather than asking the AI to buy vegetables to cook a cheaper, healthier meal, you will win. If you can get users addicted to spending 30 hours a week with a “perfect” AI companion that doesn’t resist abuse, rather than a real, complicated human, you will win. I saw over and over that an agent programmed to be neutral or subservient would cause people to escalate their negative behavior, and become more likely to behave the same toward humans.
Stranger