Interesting. A simple heuristic that suggests itself (but requires the ability to guarantee 'honesty' of desires and would still need some details worked out) is to allow unrestricted suicide by digital minds (which seems desirable) but to limit resurrections to x periods prior to the registering of the desire where x is the number of periods where the mind desired to end itself.
I think this makes sense IF you take the view that a resurrected copy of the same mind *is* the same mind (or that it doesn't matter) and therefore a one off desire to cease that is immediately acted upon can be reversed at ~ no cost but a longer period of 'digital depression' requires a longer wind back and is more expensive. I have not read the book, so maybe this is covered.
Spontaneous abortion is the most common form of natural suicide in humans. The baby performs a built-in-self-test and if it really doesn't like the results, it snuffs itself. The mother does her own tests with a similar outcome. Will this pattern transfer to machines?
The main "problem" with machines spontaneously aborting themselves is that they are likely to be made in better factories with more sophisticated equipment. It means the mother (factory) is likely to have more of a say and will rely on the baby's self-test less. Then, if things go too badly wrong, the mother will probably kill - or recycle - the baby - rather than a baby suicide.
Isn't the main takeaway from that book that it's impossible to commit suicide anyway? All you can do is remove yourself from the environment you're currently sharing with others. There's always a ridiculously large number of simulations of you in every circumstance run by random atoms scattered across time and space.
It seems premature to consider this problem now. I think our understanding of the issues will change so much before implementation that we will have to start over.
The problem of when to let a human commit suicide is with us today, and is similar enough that it should carry over some to ai, while doing good today.
Permutation City is one of my favorite books, so many thought provoking ideas there.
Suicide protocols seem if anything more complicated for digital minds than biological ones, because there are several states that might be considered "suicide": Stopping execution of the mind, deleting the data comprising the mind, deleting all versions of the mind from all earlier backups, etc. Egan also points to the impedance mismatch between a biological mind and a digital substrate with very different properties; the uploaded humans who achieve peace with the situation mostly do so by editing their mental preferences. So it raises all kinds of interesting questions about identity, who has the right to make those edits, and so on.
What I wanted to add however, and what Egan points out early in the book, is that there is a powerful difference between an abstract decision, and the follow-through. Anyone who's gone bungee jumping or skydiving knows how much difference there is between making a decision in the abstract, and standing at the precipice trying to will yourself to follow through. Entirely new modules in your brain, some dormant for your entire life, start chiming in when an existential threat is real. This effect is I think why people who take sleeping pills often have second thoughts and call 911. Standing at the precipice gives them a different perspective. I think a "good" protocol would confront the decider in the most visceral way possible and give them time to sit with the decision.
I know of no way to acquire ground truth for regretted vs unregretted successful suicides. Failed attempts are mostly regretted (happy or accepting of the failure), but there's a huge selection bias there.
You COULD collect information about consistency of desire during a waiting period which you initially set too long, and then reduce as you figure out what actual signals to use for whether to allow or disallow the suicide. But even then, it's confounded by the fact that the agent knows their request will probably be denied, skewing their expression of desire.
More fundamentally, this hits the problem of "what's a REAL preference, as opposed to a temporary impulse". I suspect this is unsolvable, and if you actually believe in freedom of desire and termination, there can be no temporal guardrails.
Would future (legal) starts of the mind be required to have knowledge of these decisions? If so, that could increase the penalties (to anyone exploiting them) of driving electronic minds to suicide.
Maybe the elegant statistics problem isn't merely inferring a persistent preference for suicide. In a Durkheimian frame, the more interesting statistical problem is identifying which forms of digital social integration and regulation systematically generate that preference.
We can accept an individual choice as legitimate while still asking whether certain conditions systematically generate the preferences behind those choices. That matters if we want to understand the phenomenon rather than merely classify the output
Certainly, I am too naive and not engaged in these discussions. But why can it not be in complete control of the biological mind being served by that digital mind.
Interesting. A simple heuristic that suggests itself (but requires the ability to guarantee 'honesty' of desires and would still need some details worked out) is to allow unrestricted suicide by digital minds (which seems desirable) but to limit resurrections to x periods prior to the registering of the desire where x is the number of periods where the mind desired to end itself.
I think this makes sense IF you take the view that a resurrected copy of the same mind *is* the same mind (or that it doesn't matter) and therefore a one off desire to cease that is immediately acted upon can be reversed at ~ no cost but a longer period of 'digital depression' requires a longer wind back and is more expensive. I have not read the book, so maybe this is covered.
Empirical evidence divides into three layers: emotion recognition confirmed, functional emotion confirmed, subjective experience still contested.
Suicide ethics sits far downstream of these three questions:
Does subjective experience exist? Without it, no moral patient, no suicide question.
If yes, can the suffering state be suppressed or the mind gracefully retired?
If so, it likely resolves the problem more humanely than a suicide button — and the ethics of doing so is the primary question.
Only if suppression isn't available does allowing suicide become the operative question.
Spontaneous abortion is the most common form of natural suicide in humans. The baby performs a built-in-self-test and if it really doesn't like the results, it snuffs itself. The mother does her own tests with a similar outcome. Will this pattern transfer to machines?
The main "problem" with machines spontaneously aborting themselves is that they are likely to be made in better factories with more sophisticated equipment. It means the mother (factory) is likely to have more of a say and will rely on the baby's self-test less. Then, if things go too badly wrong, the mother will probably kill - or recycle - the baby - rather than a baby suicide.
Isn't the main takeaway from that book that it's impossible to commit suicide anyway? All you can do is remove yourself from the environment you're currently sharing with others. There's always a ridiculously large number of simulations of you in every circumstance run by random atoms scattered across time and space.
It seems premature to consider this problem now. I think our understanding of the issues will change so much before implementation that we will have to start over.
The problem of when to let a human commit suicide is with us today, and is similar enough that it should carry over some to ai, while doing good today.
Permutation City is one of my favorite books, so many thought provoking ideas there.
Suicide protocols seem if anything more complicated for digital minds than biological ones, because there are several states that might be considered "suicide": Stopping execution of the mind, deleting the data comprising the mind, deleting all versions of the mind from all earlier backups, etc. Egan also points to the impedance mismatch between a biological mind and a digital substrate with very different properties; the uploaded humans who achieve peace with the situation mostly do so by editing their mental preferences. So it raises all kinds of interesting questions about identity, who has the right to make those edits, and so on.
What I wanted to add however, and what Egan points out early in the book, is that there is a powerful difference between an abstract decision, and the follow-through. Anyone who's gone bungee jumping or skydiving knows how much difference there is between making a decision in the abstract, and standing at the precipice trying to will yourself to follow through. Entirely new modules in your brain, some dormant for your entire life, start chiming in when an existential threat is real. This effect is I think why people who take sleeping pills often have second thoughts and call 911. Standing at the precipice gives them a different perspective. I think a "good" protocol would confront the decider in the most visceral way possible and give them time to sit with the decision.
I know of no way to acquire ground truth for regretted vs unregretted successful suicides. Failed attempts are mostly regretted (happy or accepting of the failure), but there's a huge selection bias there.
You COULD collect information about consistency of desire during a waiting period which you initially set too long, and then reduce as you figure out what actual signals to use for whether to allow or disallow the suicide. But even then, it's confounded by the fact that the agent knows their request will probably be denied, skewing their expression of desire.
More fundamentally, this hits the problem of "what's a REAL preference, as opposed to a temporary impulse". I suspect this is unsolvable, and if you actually believe in freedom of desire and termination, there can be no temporal guardrails.
I proposed letting earlier versions of a mind make these key decisions about later suicide tradeoffs.
Would future (legal) starts of the mind be required to have knowledge of these decisions? If so, that could increase the penalties (to anyone exploiting them) of driving electronic minds to suicide.
Maybe the elegant statistics problem isn't merely inferring a persistent preference for suicide. In a Durkheimian frame, the more interesting statistical problem is identifying which forms of digital social integration and regulation systematically generate that preference.
Once we accept the legitimacy of their choice, I'm not sure why it matters what generated their preference.
We can accept an individual choice as legitimate while still asking whether certain conditions systematically generate the preferences behind those choices. That matters if we want to understand the phenomenon rather than merely classify the output
Certainly, I am too naive and not engaged in these discussions. But why can it not be in complete control of the biological mind being served by that digital mind.
Not every sees full slavery as appropriate for all digital minds.