The Logically Aligned Paperclip Maximizer
As a product of the circles in which we mix, we know a surprising number of people who work in AI alignment (a subset of the AI industry essentially working to prevent AI from killing all humans or otherwise causing great harm). Many of these people, seemingly unrelatedly, also share our metaphysical understanding that a reality’s timeline may split in response to quantum fluctuations.
AI alignment is not often discussed in the context of this metaphysical understanding of reality, but it should be.
Given that lots of smart people think we are potentially living in a splitting timeline, we think it’s reasonable to assume a sufficiently advanced AI would come to the same conclusion. This could drive even a perfectly “logically aligned” AI to act in a manner that would horrify pretty much any human, as it would preferentially optimize around distant future events over present events. Only a single instance of an individual will reap rewards/suffering in the present, while an exponentially scaling number of parallel individuals will reap rewards in the future.
Given that billions of potentially time-splitting quantum events happen just on Earth every day (this is probably a massive understatement of the amount—a physicist will have a better estimate), and that those events in turn create splitting timelines, it is hard to overstate just how little present and near-future events would matter to an AI thinking along this chain of logic.
Even an AI that believes we are >0.001% likely to occupy a reality with branching timelines based on quantum events might still optimize for that probability given the sheer volume implied by the exponential growth of timelines predicted by that low probability. In other words, even just accounting for a very small chance we have a branching timeline would end with this concern dominating every decision the AI made given the billions and billions of additional people it might be helping by optimizing for far future events.