Thanks, I largely agree with this, but I worry that a Type I error could be much worse than is implied by the model here.
Suppose we believe there is a sentient type of AI, and we train powerful (human or artificial) agents to maximize the welfare of things we believe experience welfare. (The agents need not be the same beings as the ostensibly-sentient AIs.) Suppose we also believe it’s easier to improve AI wellbeing than our own, either because we believe they have a higher floor or ceiling on their welfare range, or because it’s easier to make more of them, or because we believe they have happier dispositions on average.
Being in constant triage, the agents might deprioritize human or animal welfare to improve the supposed wellbeing of the AIs. This is like a paperclip maximizing problem, but with the additional issue that extremely moral people who believe the AIs are sentient might not see a problem with it and may not attempt to stop it, or may even try to help it along.
Thanks for this comment—this is an interesting concern. I suppose a key point here is that I wouldn’t expect AI well-being to be zero-sum with human or animal well-being such that we’d have to trade off resources/moral concern in the way your thought experiment suggests.
I would imagine that in a world where we (1) better understood consciousness and (2) subsequently suspected that certain AI systems were conscious and suffering in training/deployment, the key intervention would be to figure out how to yield equally performant systems that were not suffering (either not conscious, OR conscious + thriving). This kind of intervention seems different in kind to me, from, say, attempting to globally revolutionize farming practices in order to minimize animal-related s-risks.
I personally view the problem of ending up in the optimal quadrant as something more akin to getting right the initial conditions of an advanced AI rather than as something that would require deep and sustained intervention after the fact, which is why I might have a relatively more optimistic estimate the EV of the Type I error.
Thanks, I largely agree with this, but I worry that a Type I error could be much worse than is implied by the model here.
Suppose we believe there is a sentient type of AI, and we train powerful (human or artificial) agents to maximize the welfare of things we believe experience welfare. (The agents need not be the same beings as the ostensibly-sentient AIs.) Suppose we also believe it’s easier to improve AI wellbeing than our own, either because we believe they have a higher floor or ceiling on their welfare range, or because it’s easier to make more of them, or because we believe they have happier dispositions on average.
Being in constant triage, the agents might deprioritize human or animal welfare to improve the supposed wellbeing of the AIs. This is like a paperclip maximizing problem, but with the additional issue that extremely moral people who believe the AIs are sentient might not see a problem with it and may not attempt to stop it, or may even try to help it along.
Thanks for this comment—this is an interesting concern. I suppose a key point here is that I wouldn’t expect AI well-being to be zero-sum with human or animal well-being such that we’d have to trade off resources/moral concern in the way your thought experiment suggests.
I would imagine that in a world where we (1) better understood consciousness and (2) subsequently suspected that certain AI systems were conscious and suffering in training/deployment, the key intervention would be to figure out how to yield equally performant systems that were not suffering (either not conscious, OR conscious + thriving). This kind of intervention seems different in kind to me, from, say, attempting to globally revolutionize farming practices in order to minimize animal-related s-risks.
I personally view the problem of ending up in the optimal quadrant as something more akin to getting right the initial conditions of an advanced AI rather than as something that would require deep and sustained intervention after the fact, which is why I might have a relatively more optimistic estimate the EV of the Type I error.