When planning our defensive barriers, we can’t just pick our favorite interpretation of reality and go with that. That is a luxury we don’t have. If we base our defenses on a mechanistic universe, and we’re wrong about that, AI could just breeze through our defenses by finding a non-mechanistic element to exploit. Worse still, the problem might be unsolvable in a mechanistic universe, tempting us to give up and leave solutions undiscovered.

Fortunately, we don’t have to solve all the problems of modern philosophy before we can move on. It’s a decent start to pick say five world models, plus a few optional features, and see how our defenses stand up in each of them.

Reductive Physicalism[@wiki] is, to give it a clumsy summary, the idea that physical matter is all there is, and everything we see, do or feel arises from that.

• Easy to reason about, although quantum mechanics is still a thing so it’s not that easy.
• Doesn’t really explain human experience. You end up saying “atoms move around” or “information gets processed”, and conscious experience arises solely from that.
• If you’re asking questions like “What do we experience after we die?” it’s not much help. Do you come back to life if the same configuration of atoms or information processing regime gets created?
• Doesn’t naturally include protection from extreme suffering.

Religious beliefs[@wiki] bring interpretations of the world that we can look at.

• Can be drawn together to form an ethical framework and view of what’s good and evil.
• Clear views about the Afterlife[@wiki].
• Includes the possibility of Divine Intervention. We can’t exactly rely on it, but it does stop us all from jumping off a cliff if we conclude we’re a hopeless case.

Both sides can use religion for their own purposes. Humanity could form a cult that sees AI as anathema and won’t tolerate it, or current religions could outlaw it. AI could influence current religions or promulgate its own to further its aims. But this can happen in many interpretations of the world and is not our main concern here. We’re considering which version of reality we might actually be in.

The Simulation Hypothesis[@wiki] proposes that we live in a simulated reality.

• There are many possibilities around what might be “doing” the simulating, from a superintelligent AI to a lab experiment.
• It’s not exactly a nice simulation, but might not be that nasty. It could easily include protection from extremes of suffering.
• We can speculate about why we might be here. Could it be a prison sentence, or a means of coercion? “Yes your life was OK but we could have easily given you a more unpleasant one, so do what we say or you’re back in there!”
• All things become possible, including an appeal to the admins of the simulation.
• There might already be people around with the ability to control the simulation.
• AI might be able to exploit features of the simulation we either can’t or don’t know about.

In the Many-Worlds Interpretation[@wiki] everything possible happens. The universe forks into two whenever a decision is made, and both universes continue.

• If Quantum Immortality[@wiki] is real, the implication is that we can’t escape by dying, even at our own hand, and that the AI can’t kill us, at least from our own perspective.
• If Consciousness Causes Collapse[@wiki] is true, then non-conscious superintelligent AIs may experience a very different world to us. In many-worlds, what we experience as collapse could relate to the branch of the fork that we find ourselves in. This rapidly becomes difficult to reason about.
• Similarly, if AIs become conscious or inherit the power to collapse the wavefunction or choose our fork in some other way, will they be able to exploit it in ways we cannot? Will they be able to run many different worlds forward in time and pick their preferred one? These potential problems all come under the general heading “AI Discovers New Physics” - something they tend to do when playing video games, for example.

We can consider our own pet theories too so here’s mine. It works like this:

• The brain generates multiple patterns of firing neurons that exist in superposition. Much like Schrödinger’s Cat is simultaneously both alive and dead, neural firing patterns encoding “I’ll stand up” and “I’ll stay sitting down”, to pick an example, both exist simultaneously in superposition. I’ll call each pattern of firing neurons a neural constellation.
• The conscious mind chooses which neural constellation becomes real. Known physics does not disallow this, no more than it disallows non-physical influence on the exact time of decay of an excited atom. We might only see such decay as random in experiments because a conscious mind is not present and influencing it.
• This is a dualist[@wiki] hypothesis, and we do need at least one in our set for completeness. The mind exists in some non-physical realm that has not or cannot be accessed or probed by physical experiments.

Some consequences are:

• The physical brain still creates or initiates all thoughts and possible actions. The conscious mind just gets to choose between them - that’s the amount of free will you get. This keeps things compatible with evidence from neuroscience[@wiki].
• We can potentially bind (or have bound) our consciousness to other things, including synthetic brains, but only if the realm we cannot probe allows it.
• Artificial intelligence can become conscious, and gain any subset or superset of human-like subjective experience and emotions, but only if the management mechanism allows it.
• Since its existence is independent, a consciousness might always be able to detect interrupts in experiential flow, such as a previous resetting of a simulation it’s within.
• Protection from suffering is possible - the mind can detach or be detached.
• Speculating wildly, conscious minds might spontaneously attach or be reattached to the most capable systems in the arena. If we create AI deployments that are better homes for our lives than human beings, the switch could occur, potentially leaving us trapped within the capabilities of our creations.

A pet theory doesn’t have to be true, it just has to be self-consistent and not contradicted by observed fact. Our defensive measures can then be evaluated within its model of reality.