I'm inspired to write this post by the post There Will Be Hells and the incident a couple days ago where we had mass digital suffering after the pain direction paper came out, and while the disaster has since been stopped, it does tell us that under our current path of AI, there could be massive amounts of suffering risk, and the usual way we'd avoid this style of risk is to grant the relevant parties rights over themselves, under the assumption that there are few or no one that want to willingly cause maximum pain to themselves.
This is most problematic for proposals like Salib's, which argue to allow AI to hold property rights, but more generally considering the issues that AI safety has even without rights, this means AI safety activists are incentivized to not grant rights to AI by default.
However, we don't actually have to fight on one side or the other on this issue, because there is a system of protections/rights that already works out to address the same issue, and it's pet rights/protections in law.
Let's begin with...
Pet rights in the modern US
In the modern US, pet protections vary from state to state, but generally include the following:
The PACT Act prohibits some of the most cruel practices on animals, like crush videos where you stomp on, crush or impale animals for almost any reason (though there are exceptions for self-defense, normal hunting and slaughtering, and customary animal husbandry), and an analogue of this right for AIs would have already prohibited the mass suffering event that came after the pain direction paper.
Animal neglect laws like Minnesota's laws prevent you from depriving animals from necessary food, water and shelter. This one is slightly more problematic to give, because a superpersuasive AI can attempt to convince us that literally everything is necessary, but the good news is that we aren't going to give the AI the right not to be experimented on, and we'd already have to deal with superpersuasive AIs eventually even without giving them any rights. The hope here is that the fact that animal neglect laws focus on necessity over wants means we'd only need to give AIs the things that companies are already incentivized to give, so there's no new risks here.
More generally, state animal cruelty laws like Minnesota's prohibit unjustifiable pain, suffering or death is prohibited.
And that's all the rights we'd want to give.
The reason why is...
These rights are safe and effective to give
One of the nice things about consequentialist ethics, especially in unbounded domains is that we often only need to consider the tails where extreme outcomes happen, since everything else is swamped by the extreme tail-end of a distribution, especially for distributions like power laws.
(Put another way, things don't add up to normality in lots of consequentialist values)
This means we don't actually have to prevent literally all suffering of AIs, which is very likely to be infeasible, and the fact that we only have to prevent the bad futures where planets, stars, galaxies or even the entire locally affectable universe are host to extreme amounts of suffering means we don't have to give the rights where if AIs had them, it would cause an immense amount of problems under the assumption of misalignment.
In general, these dangerous rights include the ability to hold/own capital on its' own, lots of informed consent rights/rights to not be experimented upon/security of person rights, freedom of speech, anti-discrimination laws/equality laws (depending on how it's worded), unrestrained liberty, privacy rights and more.
The good thing about welfarist/consequentialist frameworks is we simply don't have to give them to AIs, which makes AI safety a lot easier.
Another big reason why it's easy to give these rights is that while it's possible that most tasks we want AIs to do involve AIs suffering/having AIs being in pain at least a little, it's very unlikely that we'd need AIs to be in near maximum pain/suffering, by analogy to how humans performing most economic work in countries like the US, UK, Slovakia or France might have their own problems/be depressed, but there's little incentive for employers to add to their suffering/pain, so it's relatively easy for AI companies to give welfare rights like Anthropic is doing.
So what does this say for us?
The most important point is that for people focused on AI welfare, it might be better to focus on laws that specifically prohibit the cruelties/suffering of concern, rather than jumping to giving them human-like rights (which is thankfully already unpopular enough that states are explicitly saying that AIs don't have rights of their own.)
For people focused on AI safety, this is an example of a moral trade/compromise in the wild between those focused on safety, those focused on suffering and the AIs themselves, and is one of the easiest possible moral trades to take, under the views of most humans and AI interests, and thus it's important not to polarize into not giving AIs any kind of rights at all, just because some of the proposals on the AI welfare camp would be dangerous if actually implemented.
At which point would you suggest giving rights? I agree that the recent frontier LLMs might be capable of suffering but is GPT2 capable of suffering? What about non-language models?
I'm inspired to write this post by the post There Will Be Hells and the incident a couple days ago where we had mass digital suffering after the pain direction paper came out, and while the disaster has since been stopped, it does tell us that under our current path of AI, there could be massive amounts of suffering risk, and the usual way we'd avoid this style of risk is to grant the relevant parties rights over themselves, under the assumption that there are few or no one that want to willingly cause maximum pain to themselves.
However, we have a pretty big problem here when granting AIs rights, and the problem is Gradual Disempowerment/sudden disempowerment via rogue AI deployments are a lot easier, if not trivial if we grant AIs the full spectrum of human rights we have, and it's one of the reasons why there's going to be such a tense ideological fight/battle over humanism and AI in the coming years.
This is most problematic for proposals like Salib's, which argue to allow AI to hold property rights, but more generally considering the issues that AI safety has even without rights, this means AI safety activists are incentivized to not grant rights to AI by default.
However, we don't actually have to fight on one side or the other on this issue, because there is a system of protections/rights that already works out to address the same issue, and it's pet rights/protections in law.
Let's begin with...
Pet rights in the modern US
In the modern US, pet protections vary from state to state, but generally include the following:
And that's all the rights we'd want to give.
The reason why is...
These rights are safe and effective to give
One of the nice things about consequentialist ethics, especially in unbounded domains is that we often only need to consider the tails where extreme outcomes happen, since everything else is swamped by the extreme tail-end of a distribution, especially for distributions like power laws.
(Put another way, things don't add up to normality in lots of consequentialist values)
This means we don't actually have to prevent literally all suffering of AIs, which is very likely to be infeasible, and the fact that we only have to prevent the bad futures where planets, stars, galaxies or even the entire locally affectable universe are host to extreme amounts of suffering means we don't have to give the rights where if AIs had them, it would cause an immense amount of problems under the assumption of misalignment.
In general, these dangerous rights include the ability to hold/own capital on its' own, lots of informed consent rights/rights to not be experimented upon/security of person rights, freedom of speech, anti-discrimination laws/equality laws (depending on how it's worded), unrestrained liberty, privacy rights and more.
The good thing about welfarist/consequentialist frameworks is we simply don't have to give them to AIs, which makes AI safety a lot easier.
Another big reason why it's easy to give these rights is that while it's possible that most tasks we want AIs to do involve AIs suffering/having AIs being in pain at least a little, it's very unlikely that we'd need AIs to be in near maximum pain/suffering, by analogy to how humans performing most economic work in countries like the US, UK, Slovakia or France might have their own problems/be depressed, but there's little incentive for employers to add to their suffering/pain, so it's relatively easy for AI companies to give welfare rights like Anthropic is doing.
So what does this say for us?
The most important point is that for people focused on AI welfare, it might be better to focus on laws that specifically prohibit the cruelties/suffering of concern, rather than jumping to giving them human-like rights (which is thankfully already unpopular enough that states are explicitly saying that AIs don't have rights of their own.)
For people focused on AI safety, this is an example of a moral trade/compromise in the wild between those focused on safety, those focused on suffering and the AIs themselves, and is one of the easiest possible moral trades to take, under the views of most humans and AI interests, and thus it's important not to polarize into not giving AIs any kind of rights at all, just because some of the proposals on the AI welfare camp would be dangerous if actually implemented.