Dissolving the Problem Of Evil

July 23, 2022

The problem of evil has a hidden assumption: that a limited observer and an omniscient evaluator must reach the same judgment about what constitutes good action. They can’t. Provably. The divergence appears the moment you add even one unit of additional information.

The argument states:

It is making an argument which is claiming that set x is guaranteed to be equal to subset(x). We can know there is a full x because omniscience is introduced. We know there is a subset(x) because we aren't omniscient.

It declares a guaranteed equality of evaluation function over all possible data subsets of an omniscient entity. Humans evaluate f(subset(x)). The omniscient evaluates f(x).

If f(subset(x)) = f(x) in a general way then subsets are equal to sets. If the problem of evil is a valid argument, then it must be the case that 0 = 1. It isn't a valid argument. See further below for a simulation demonstrating that limited information produces the observation of evil even when evil does not exist.

Less Informed Agent

Score:

Moves:

Average Score Per Move:

More Informed Agent

Score:

Moves:

Average Score Per Move:

I've hidden part of the board above which calls attention to the way more information lets you gain utility despite that utility not being visible. I'm also only showing one simulation, because it should draw attention to the statistical nature of the issue. Some will watch this and conclude from their sample a falsehood - that the two agents were of equal ability. They aren't.

A somewhat interesting caveat of great import only to certain personalities is the that the average game length for the more observant agent is higher. This has radical implications for how an intelligent agent ought to be made.

P1. If an omnipotent, omnibenevolent and omniscient god exists, then evil does not.

P2. There is evil in the world.

C1. Therefore, an omnipotent, omnibenevolent and omniscient god does not exist.

P1 is where the argument breaks. It assumes that an omniscient evaluator and a limited observer, applying the same value function to different datasets, must agree on what is good. They don’t have to. Below is a simulation that shows why.

Two agents share the same reward function. One agent can see one space ahead. The other can see two. One agent can see one space in front of it. The other is ever so slightly expanded in terms of what it can see. It can see one additional unit. Observe that despite sharing the same function for valuing situations, the two agents can disagree as to what action is good to take. The one space agent will never explore areas in which there is a negative reward, but the two space agent will do things that the one space agent considers evil, because it could make better decisions on account of additional information.

What we’re seeing here is that when you evaluate the problem of evil from first principle, using the part of the mind that is comfortable with thinking about things slowly and in depth, rather than using the intuitive part of the mind which short-circuits evaluation, the argument reveals itself to be fundamentally broken. Not only is it not true that omniscience results in a lack of evil, it creates the conditions for us to observe someone else doing evil when in fact they are doing good.

This happens well before you get to omniscience. It happens the moment you add even one more unit of capacity to observe what is happening.

This should be obvious. If we have the same function, but we apply it to two different datasets, clearly the result can differ. Each additional unit of information just increases the likelihood of divergence.

I’ve taken to calling this error the horizon effect, because when I see it happen it appears to happen on account of people having an information frontier which they haven’t explored. They don’t recognize that the dataset they evaluate isn’t actually the only dataset present. They don’t extend their thinking beyond the horizon, but keep it fixed on what is near at hand.

It makes sense that people do this.

We have a limit to how much time we have. We have to use that time wisely. As a consequence, strategic laziness is essential. Real wisdom recommends diligence, but insight tells us that strategic laziness and strategic dilligence are the same thing: a choice about where to apply our time in order to be effective. So people do it and sometimes they make the inevitable mistakes.

We can’t explore every frontier. We can’t push every horizon outward. We simply don’t have the time to do so. If just one step of additional information is enough to cause differing evaluation functions, it is quite obvious that extending it an infinite amount opens up the opportunity to even more divergence.

This might feel like sloppy reasoning to someone who really likes the problem of evil. There are, after all, genuinely terrible things. Innocent children die horrible deaths.

But the argument isn’t about whether evil exists from our perspective. It is about whether observing evil is sufficient to conclude that an omniscient agent is not acting toward good. The simulation shows it isn’t. More knowledge can lead to optimal decisions that appear sub-optimal to a less informed observer. The two-space agent accepts short-term negative rewards because it can see that they lead to higher total return.

The practical takeaway extends beyond theology: prefer learning to judgment. When someone with more information makes a decision that looks wrong to you, the gap might be in your information, not their values.

The code to generate the visualization, imports not included: