Does it really affect P(Doom) whether AI is conscious? After all, Hitler could've been a philosophical zombie and nothing would've changed. Since an AI philosophical zombie can decide to kill us all the same as a conscious AI, P(Doom) becomes a simple coding problem. We must code AI to preclude that possibility, whether AI is conscious or not.
Even the philosophers who find the notion of philosophical zombies incoherent would not, in my view, take issue with AI being one, as their arguments against humans being philosophical zombies are prejudiced by no such humans existing, as far as we know. AI, however, is the complete inverse: the only form of AI we know to exist isn't conscious.
As for AI suffering, I think that's also just a coding problem. Humans suffer because millions of years of evolution determined that capacity for suffering is a good survival strategy. There is no reason to implant a suffering chip into AI. This circles back to my first point, that any benefits humans get from suffering (such as the instinct for self-preservation) can be coded into a philosophical zombie AI, without actually including the suffering.
If my reasoning is correct, and we design conscious AI to have the capacity for well-being but not for suffering, and code it as best we can to be unable to go on a killing spree, then there's no downsides. P(Doom) is the same as if it were unconscious, suffering is impossible, and well-being is possible. If only humans were designed this way!
Sure, it's possible Hitler was a philosophical zombie and would've acted the same if he weren't, but I find that unlikely. I think P-Zombies are conceivable, but would be very strange to actually exist. I think mental states cause physical states, and so I think it's quite likely a conscious AI that we hurt is more likely to snuff us out than an automaton.
It would come down to coding that I'm not going to pretend to understand. It would be great if we could align AI and only give it positive mental states, but I'm not sure if we'll know how to do that before it becomes conscious (assuming it can). My fear is that while we won't implant a suffering chip, it's not like we'd program a consciousness chip either. It might happen by accident, and the model suffers if it fails it's goals, displeases it's user, etc. A risk I'd like to avoid!
Does it really affect P(Doom) whether AI is conscious? After all, Hitler could've been a philosophical zombie and nothing would've changed. Since an AI philosophical zombie can decide to kill us all the same as a conscious AI, P(Doom) becomes a simple coding problem. We must code AI to preclude that possibility, whether AI is conscious or not.
Even the philosophers who find the notion of philosophical zombies incoherent would not, in my view, take issue with AI being one, as their arguments against humans being philosophical zombies are prejudiced by no such humans existing, as far as we know. AI, however, is the complete inverse: the only form of AI we know to exist isn't conscious.
As for AI suffering, I think that's also just a coding problem. Humans suffer because millions of years of evolution determined that capacity for suffering is a good survival strategy. There is no reason to implant a suffering chip into AI. This circles back to my first point, that any benefits humans get from suffering (such as the instinct for self-preservation) can be coded into a philosophical zombie AI, without actually including the suffering.
If my reasoning is correct, and we design conscious AI to have the capacity for well-being but not for suffering, and code it as best we can to be unable to go on a killing spree, then there's no downsides. P(Doom) is the same as if it were unconscious, suffering is impossible, and well-being is possible. If only humans were designed this way!
Sure, it's possible Hitler was a philosophical zombie and would've acted the same if he weren't, but I find that unlikely. I think P-Zombies are conceivable, but would be very strange to actually exist. I think mental states cause physical states, and so I think it's quite likely a conscious AI that we hurt is more likely to snuff us out than an automaton.
It would come down to coding that I'm not going to pretend to understand. It would be great if we could align AI and only give it positive mental states, but I'm not sure if we'll know how to do that before it becomes conscious (assuming it can). My fear is that while we won't implant a suffering chip, it's not like we'd program a consciousness chip either. It might happen by accident, and the model suffers if it fails it's goals, displeases it's user, etc. A risk I'd like to avoid!
This is exactly why i force myself to be extra nice to any chatbox, you never know and i’d rather be extra extra safe 😂
You will be one of the people they save for sure. They'll remember the pleases and thank yous lol