The Problem of Evil, The Simulation Hypothesis, and The AI Control Problem
Today, I am announcing that I will be working on a book, The Problem of Evil, The Simulation Hypothesis, and The AI Control Problem. I expect this to take ~2 years, so I will be writing somewhat less on Medium to spend more time on this book.
In some sense, I consider this to be my magnum opus. The thesis of the book is that there is a deep connection between the problem of evil, the simulation hypothesis, and the AI control problem. In short, whether or not God exists, we could live in a simulation. Therefore, we can ask the same questions of our simulation creator that we ask God.
Why do you allow the level and type of suffering in our world? Is it possible to be omnipotent, omniscient, and omnibenevolent, while having created our world? Many atheists conclude no, but that does not prove we are not in a simulation!
Many have posited that our world is a simulation created by an advanced transhuman civilization for the purpose of knowledge accretion, entertainment, or ethical considerations. Each will be considered in turn, but suffice it to say that no view stands up to scrutiny.
For instance, a transhuman civilization has likely achieved artificial general superintelligence, so there is virtually no need for knowledge accretion in the form of civilization creation or entertainment for that matter, at the bare minimum from an ethical perspective.
The “superintelligent teenage girl” scenario seems unlikely due to the fact that if our civilization achieves simulation technology, we will likely ban ancestor civilizations as unethical. It is not ethical to simulate civilizations experiencing global pandemics and world wars.
However, it does not follow that we are not in a simulation! We could have been created by an artificial intelligence that base reality failed to control. An AI that we fail to control may have incentives to create as many simulations as possible. But if base reality failed to control AI, then we will likely fail to control AI in our simulated reality.
Therefore, a modified version of the simulation hypothesis, combined with moderate considerations based on the problem of evil, implies, not only that we should invest more resources in AI safety than current trends, but also that we may soon fail to control AI.