Digital Mind Suicide

Greg Egan’s 1994 novel Permutation City depicts digital minds having a right to suicide; if they push a particular internal button, they end. And this does seem like a robust way to avoid many horrible scenarios. (I also discuss this idea in AgeOfEm.) But the novel also depicts such buttons sometimes being illegally disabled, and sometimes others frequently restarting recent copies from before a suicide request. Both of which undermine the policy.

It could make sense to wait for a confirmation after a delay, to check if a suicide request was due to a temporary mood fluctuation. It could also make sense to disable new starts from archived copies from a period before the suicide request. But such policies seem to have many free parameters; how could we set those parameters in a principled way?

One approach is to let each digital mind set their own parameters re such policies. But if they could do this at any time, they might use this power to escape these policies. So it could make sense to only allow these policies to be reset with a substantial delay. But now that adds another policy parameter.

It seems to me that we want to think about this in terms of statistical inference. The state that we want to infer is a stated desire to commit suicide that is continued over a sufficiently long time period. We want to infer this state about every point on an entire tree of digital minds descendancy histories. And our main data for inferring this is explicit choices to request suicide, and confirmations of such requests. If for any point in this tree we can infer a high enough (yet another parameter) chance of continued willingness to confirm suicide, that digital copy should end.

It seems to me that this is a nice elegant stat problem, for which stat experts should be able to come up with some simple yet powerful inference rules. Of course given huge datasets it would be possible to generate far more complex and accurate models with their matching inference rules. And maybe digital minds should be allowed to choose such more complex inference rules. But let’s first figure out some simple ones.

A key issue is how to ensure that such inference rules get access to sufficient data on the local digital mind descendancy tree to say which copies should end. That will depend on privacy rules on digital mind descendant relations and suicide requests. It could make sense to give a special cheap suicide monitoring process access to more such info than usual, info only used to apply suicide inference rules. Even today we prioritize suicide relevant info info in privacy rules regarding counselors.