The punishment dilemma

The root of the "dilemma" just seems to be that we find it unlikely that a trustworthy model would really say such a thing. Human brains have trouble granting false premises, even hypothetical ones.

So yes, I would cut all punishments to zero if a sufficiently trustworthy model told me it would be for the best. And if a sufficiently trustworthy model told me I could fly by flapping my arms and quacking, I'd be outside pretending to be a duck.

You're asking, then, whether I would trade off the satisfaction of having other tribe-members help me punish the one who hurt me, against the probability of being hurt? In the specific circumstances you have set up, I absolutely would. The reason is that, if there's no punishment for murder, there is presumably likewise no punishment for retaliation in kind. So I don't need my tribe to enact any punishment, I can do it myself, gaining even more satisfaction. This really seems like the best of both worlds: You get to take personal vengeance on anyone who wrongs you, with no dilution through 'official' coalitions; while the probability of actually being wronged goes down.

People want both (as separate values) safety from crime and retribution for crime. In this thought experiment, zero punishment is stipulated to maximize safety from crime, but it simultaneously minimizes retribution. If marginal [value of] retribution at zero punishment is lower, then zero punishment is optimal, otherwise a non-zero amount of punishment is better overall. Alternatively, if the only available forms of punishment are such that the aspect of (undesirable) hurting of criminals outweighs the aspect of (desirable) retribution for the crime, then zero punishment is automatically optimal in this thought experiment.

crime rates anti-correlated with harshness of punishments imposed on criminals.

This sentence appears to have a typo; if crime anticorrelates with punishment then more punishment equals less crime.

So it helps people in general by decreasing crime, and helps criminals by not hurting them? Sounds like win-win to me!

Punishment hurts people. It's terminally bad. I'm only okay with it because it keeps other people from being hurt, so its effect is net good.

I am suspicious of thought experiments involving especially implausible situations, but I would say that if circumstances were as you described, the demobilization would be morally required.

I don't know whether the value I get from revenge is greater than the value I get from decreased danger, but I am almost sure that would depend on exact numbers and there would be a threshold T of probability for each crime. So, for example, let P = present law and H = hypothetical law which doesn't punish murder and p(M) = murder rate; I would prefer H to P iff p(M | H) is lower than T(murder). I am not certain what T(murder) actually is, but it may also depend on p(M | P). To say at least something, I would prefer P to H if p(M | P) - p(M | H) < 10^(-8) , and I would prefer H to P if p(M | H) < 10^(-8).

If you want to survey about values attributed to retribution, you could probably do it more directly without assuming an inplausible scenario. Just: "Would you choose to do action X, if the consequences were (1) reduction of crime rate by amount Y, and simultaneously (2) no punishments for crimes?" There is no need to assume that (2) is the cause of (1); doing so primes people to concentrate on the implausibility instead of the core of your question.

A reformulated remotely plausible version of the thought experiment: I am a rich experimental psychologist / experimental ethicist / sociopath. You are dictator of a city state with 1 million inhabitants. I offer you to install a super-powerful automatic security system if and only if you agree to abolish all punishments. The automatic system covers your entire city by surveillance cameras and detects all crimes. Moreover, there are robots which can prevent 99% of crimes (private retributions included) from happening, and experiments show that the increase of attempted crimes in the absence of punishment is only by factor of 10. This means that the overall number of committed crimes decreases 10 times. Would you accept my offer?

Least convenient possible world, of course: you have absolute power, so no real chance to be overthrown because of making unpopular decisions; the security system is safe and destroys all records after they are of no use for the crime-preventing robots, so no disutility associated with loss of privacy; etc.

One additional assumption: there is no significant target shift.

Armed robbery for taking week's profit from ten businessmen may be less of a strike to all of them in sum than armed robbery taking months' salary from a minimal-wage-working single mother. It would count for a greater crime rate, though...

If all this were true and if the model held around zero, then doing the thing that minimizes suffering would be good.

That doesn't mean we should not check if other forms of punishment would drop crime rate further.

Of course, the model would have to include a lot of stability checks. With 1¢ fine for every crime, people would use a lot of interesting weapons for self-defense in broad sense (including defending good neighbour's property from theft by force just in case). People wanting retribution would do the retribution themselves... Some other interesting effects would occur...

If that worked out to lower aggregate crime rate for each victim group - it would be nice...

Your assumptions are not just unrealistic, they do not appear to be self-consistent, which is much worse.

"Crime is the breach of rules or laws for which some governing authority (via mechanisms such as legal systems) can ultimately prescribe a conviction." If you abolish punishment, you abolish crime, however silly it sounds. There is no longer a distinction between immoral and illegal. The "crime rate" would be trivially zero. The "moral offense" rate would be hard to calculate without restoring some gradation of immorality, otherwise being rude to a person has the same weight as killing them.

I see no way to create this gradation other than by assigning different punishment to different offenses, in which case we are back to where we started.

A more consistent model would be "what if there is a maximum level of punishment beyond which there is not further reduction in crime rates, and this level is entirely too low to the victim's liking?" When stated this way, the answer is obvious: most people would adjust their expectation of a just punishment to fit that prescribed by the law. This has happened over the ages in nearly every society already (the adjustment, not the punishment optimization), in one direction or another.

For example, capital punishment used to be dispensed rather freely not that long ago. On the other hand, swindling was not a crime until rather recently, and often still isn't. Societies adjust to what the law says, and change the law when there is enough support, however imperfectly.

Given this self-adjustment of expectations of a just punishment, all that remains from your question is "how to find the optimal level of punishment for a given offense?" A zero-punishment offense would be considered an immoral act, all the rest would be forms of infractions/civil offences/crimes.

I would be a one man counter example in such a society. I doubt I'm unique. Your thought experiment assumes no one that is like me.