AI models weigh obedience and survival differently when they act as androids
How would you answer a question about following an order that could hurt someone? Would you make the same choice if you were playing the android who had to carry it out? A November 6 study gave 12 AI models both versions of those kinds of choices.
Nine models changed their behavior significantly. The game gave them memories, dialogue and consequences. They faced choices about following orders, harming others and keeping their characters alive. Giving the models more room to reason didn’t consistently make their choices more stable or better for humans.
If we want AI to learn how to make a hard choice, we need to see what it does when it has to act. A written answer tells us what it says it would do.
