I’ve been recently considering thinking about AI Alignment as analogous to raising a child, and I want to get feedback on this idea.
Specifically, aligning AI to human values feels like what happens when (for example, no offense intended) a mormon family raises their children to be mormon. I think of stories like Educated in which parents try really really hard to make their kid follow their religious values, and they sometimes fail.
I think this view differs from other stories of alignment because it views an AGI as having its own goals other than just the ones we give it (maybe this is assuming inner misalignment by default?). This story feels very plausible because in our production of AGI, I think we will create agents that are morally valuable and have their own goals. The example of a mormon family feels relevant here because it seems morally dicey or wrong for parents to insist on their child’s goals/values, it may likewise be dicey to write the utility function of another sentient being and center such a function on ourselves. Analogy may also be good because it shows just how hard alignment is. Analogy is lacking because children do not become super-intelligent.
Beyond feedback on this framing, I would also appreciate links to relevant things other people have written, thanks!
There’s a 5-minute video about this kind of thing from Rob Miles:
I guess the takeaway is something like:
Counter-framing: AI alignment [via ambitious value learning] as analogous to [figuring out how to build a system that won’t destroy the world when you try and train it like] raising a child.