How the simulation hypothesis is closely related to the problem of evil, existential risk, and the AI control problem
Follow me: The Moral Economist
During Recode’s 2016 Code Conference, Elon Musk shocked the non-philosophical world by arguing that we are in a simulation. Musk claimed “Either we are going to create simulations that are indistinguishable from reality, or civilization will cease to exist.”
Therefore, he argues that we should hope we are in a simulation, since the probability we are in base reality (OG) is nearly statistically impossible (per the simulation argument), and if we were in (OG) reality, we may still face an existential hurdle (closely related to Fermi’s paradox).
The Simulation Hypothesis (SH) may seem like an innocuous modern rendition of Descartes’ demon; however, it is based on empirical considerations, not radically skeptical ones. Furthermore, SH has potential implications for existential risk that must be considered; one cannot simply dismiss it out of hand.
In particular, I argue (per the simulation argument) that denying the simulation hypothesis commits one to either 1. denying a weak form of technological determinism, 2. assuming conscious mental states cannot supervene on any substrate (strong dualism), or 3. all societies in all quantum worlds in the multiverse that reach our level of technological development fail or choose to not run simulations.
Indeed, there’s an odd connection to the ancient problem of evil. Namely, if you assume we are in a simulation, why was our world simulated? If we are in a simulation, is our creator morally praiseworthy?
Is the degree and type of suffering in our world compatible with a “good” creator? (The Problem of Evil) If the answer to this question is no, it does not follow that we are not being simulated. For example, if we are being simulated for purposes of knowledge accretion, this may have enormous consequences for existential risk.
On the other hand, self selection is relevant to our considerations. Even if there exist vast numbers of morally ambiguous or morally malevolent simulations, it does not follow that we are necessarily in one of those simulations. If our creator is morally praiseworthy or if we one day reach the technological ability to create simulations but choose not to, then we are likely not in a simulation.
I argue that societies such as ours are likely to ban conscious suffering simulations, while amoral agents (such as an AGI) may not consider the moral status of conscious beings and therefore create vast numbers of simulated worlds. This would seem to indicate that the probability we are in a simulation at all is related to the probability our creator (or lack thereof) is morally praiseworthy.
I discuss the relationship between the simulation hypothesis, the AI control problem, the problem of evil, and the vulnerable world hypothesis. In particular, I argue that whether or not we solve the control problem is related to the probability that we are in a simulation.
Finally, the vulnerable world hypothesis “constrains” possible simulations. For example, the “lazy person in their basement” can’t necessarily create simulated worlds if restricted through a global governance structure.
As to the title “Elon Musk is Right We Might be in a Simulation, Wrong That Means We’re Okay”, while many may think that the simulation hypothesis is absurd, it is far from refuted. Moreover, the simulation hypothesis being true does not mean we face less existential risk than if we were in OG reality; we could be simulated for a variety of reasons (morally praiseworthy, morally ambiguous, or morally malevolent).
Just because the world not blowing up is a necessary condition for the simulation hypothesis, it does not follow that being simulated means the world will not blow up. This article adds yet another motivation for the urgency and seriousness of solving the control problem.
Nearly 500 years ago, mathematician and philosopher Rene Descartes offered a mind altering thought experiment now known as Descartes’ demon. He asks us the counterintuitive question, how can we know anything at all? I may think I am typing on a laptop, but how can I possibly know that there isn’t an evil demon controlling my world? I mean Donald Trump is president, right?
You might debate this by saying that demons don’t exist, but that wouldn’t have satisfied a theist like Descartes. Furthermore, this example is part of a broader branch of philosophy known as radical skepticism. Bertrand Russell asks another hypothetical unrelated to God’s existence. How do we know for a fact that the world wasn’t created 5 minutes ago with all your memories just recently implanted? There would be no way to prove otherwise.
Descartes concludes we can at least know one fact. I think therefore I am. Whatever false, orange images a demon may hypothetically be projecting onto my mind, I know I at least experience something. From this, Descartes builds a rational philosophy, first by attempting to prove God’s existence, and then concluding that surely God wouldn’t allow a demon to trick his creatures; therefore, we can trust our mental faculties.
Russell disputed these arguments for the existence of God, so he was faced with the same skepticism that plagued Descartes. (Hence why his thought experiment is devoid of religious connotations.) Ultimately, his rejection of radical skepticism (RS) is surprisingly simple.
He turned RS on its head. What if we held RS to its own standard? In other words, we should be skeptical of skepticism. It is simply too high a standard for knowledge. If we live our lives thinking that everything started 5 minutes ago or a demon is controlling our thoughts or I am the only person in existence (solipsism) or I am a clone (like I believed in 5th grade after reading House of Scorpion), then we basically cease to know anything.
We must make certain minimal assumptions to allow any knowledge accretion at all. We can reject any RS argument by asking what independent reason do you have to believe in that RS argument? What independent reason do we have to believe in a controlling demon or that the world was created 5 minutes ago? For now, we have none, so we are justified in not believing in them.
If, on the other hand, you notice patterns of people watching you and find a God door at the edge of your globe like in The Truman Show, you have independent proof that your worldview is in fact a conspiracy.
The Slightly Different than Matrix Hypothesis
This all seems closely related to the simulation hypothesis, a famed thought experiment by legendary Oxford philosopher Nick Bostrom. The simulation argument (SA) asks how do we know we are not in a computer simulation?
(The simulation argument (SA) aims to prove that either 1. technological progress must end, 2. all civilizations don’t run simulations, or 3. we are in a simulation. The simulation hypothesis (SH) is the belief in 3., that we are in a simulation.)
Just like other skeptical seeming arguments, there appears to be no empirical (evidence/fact based) way to prove we are not in a simulation. Therefore, many dismiss the simulation argument on the same basis as other RS arguments. However, Bostrom’s paper is not at all an RS argument. Instead, Bostrom argues that we have empirical reasons to believe in SH.
Namely, we have already seen the incredible progress in video games, movies, graphics, and technology. If we assume any rate of progress at all, it is inevitable that we will be able to fully simulate human minds, simply based on estimating the upper bound of the number of neurons needed to simulate a human brain times the number of all humans that have ever existed, plus estimating how much computing resources it would take to do this task. Many many times over.
(Bostrom’s paper is littered with notes such as “An alternative estimate, based on the number of synapses in the brain and their firing frequency, gives a figure of ~10^16–10^17 operations per second.”)
Hence I argue SH is an empirical argument, not a theoretical one. What Bostrom’s paper proves is close to Elon’s initial contention that “either 1. we are going to create simulations that are indistinguishable from reality, or 2. civilization will cease to exist.”
Bostrom qualifies this argument by adding a third condition: or 3. we will not run simulations at all (either because simulations are not technologically possible or because we decide not to create a simulation for ethical or other reasons). In doing so, he strengthens the logical validity of SA:
1. Technological progress stops for any civilization that reaches our level of technological development,
2. No civilization runs simulations,
3. We are almost certainly in a simulation.
(Note for the philosophically inclined: If ~2., then 3. by assumption of relatively large numbers of simulations per base reality. If ~1., then 3. by assumption of substrate-independence. See section II of Bostrom’s paper.)
So, let’s look at the empirical facts and implications of SA’s three possibilities.
The Nuance of The Simulation Argument
Once we accept that technology is improving at a rate consistent with simulating many human minds, it is natural to ask how do we know we are not already in a simulation? Let’s say in our world it takes until 2100 to be able to simulate a human brain. (More plausibly, Ray Kurzweil estimates this will happen in the 2030s.) Suddenly, we can just create life. Maybe some kid on his XBox 2100 can simulate a few thousand souls in Sim’s 2100.
Can we simulate human minds?
If you were one of those souls, everything would seem the same as our world. You go to school, learn philosophy, create technology, go to the bathroom, make friends, experience life. The number of simulated beings would most likely be nearly infinitely larger than the number of humans in “base” reality (the first world to create sims).
To understand why, our XBox kid’s computing power in 2100 would most likely be much more powerful than all of the world’s computing power today, per Moore’s Law. That would make it easy to simulate entire worlds, full of hundreds, or even billions, of people.
If the Xbox 2100 isn’t powerful enough, perhaps humans in 2300 decide to use all of their world’s computers (at that point, likely the size of a planet) to simulate as many worlds as possible. (Why would they do that? Well, that’s kind of the point of my paper!) What difference does a couple hundred years make on evolutionary time scales?
There is a key assumption to make this argument possible. But it’s not what you might think. You may want to avoid this argument by retorting, “We can’t know that we will actually be able to simulate a human mind, even if we may hope that we will one day be able to do it.” The argument you probably want to make is actually “We can’t know that we will be able to simulate consciousness,” not a human mind.
This is relevant, because empirically speaking, it seems quite likely that we will one day have the computing power and knowledge of neuroscience to mathematically map neurons and synapses in a way that runs on a computer. If we don’t in our world, it’s likely because technology ceased to progress, rather than it was physically impossible.
Look at how good video games and CGI are in 2020 compared to 2000. Most sports games are approaching realism, Call of Duty is pretty impressive, and we have the ability to bring back dead actresses with CGI. We all know how fast exponential processes accelerate. Today’s deep neural nets can already run billions of calculations in a manner not entirely foreign to the human mind. We will likely have the computing power to run full brain emulations within 100 years.
And even if we can’t for whatever reason, given how close our civilization already is to creating simulations, is it really implausible that a single civilization in all of the multiverse, in all quantum time periods, reaches this level of technological proficiency? If you think this is even possible, the SH becomes relevant again. Because then it is an empirical question to which we can attach Bayesian probabilities, just like we do with the Drake Equation.
For the rest of this article, I assume the following:
[Mathematical Coherence of Brain Functions (MCBF)]
It is theoretically possible to write mathematical equation(s) that describe the microscopic (neurons) and macroscopic (interactions/constraints/emergent properties/etc.) processes of brain states, as well as their evolution over time.
Denying this amounts to a form of strong dualism. Imagine that after 10,000 years of progress in neuroscience, we still can’t describe the human mind with mathematics. Then, there are “mysterious” processes that are probably what dualist have in mind (pun intended) when they deny reductionism.
However, not all forms of dualism are like this. For example, David Chalmers is a scientific materialism, and he believes that consciousness is a fundamental variable, like electromagnetism or the speed of life. Therefore, these “weaker” forms of dualism are still non-reductionist and allow for my assumption of MCBF.
Can we simulate consciousness?
A stronger rejection that we are in a simulation comes from denying the ability to simulate consciousness. In Bostrom’s parlance, we must assume “that mental states can supervene on any of a broad class of physical substrates.”
A mind/body dualist could argue that it is impossible to create consciousness with raw physical matters, even if we could simulate all of a brain’s movements. Most dualists are also theists, and most atheists are not dualists; however, dualism need not imply theism. (See Thomas Nagel.)
Nevertheless, theists and dualists can still believe simulating consciousness is possible (there’s nothing contradictory about God allowing conscious simulations), and nearly all materialists are forced to believe that consciousness can be simulated.
Why? Because if materialism is true, we, physical beings, are conscious. It seems arbitrary from the perspective of a scientific materialist to believe that only biological molecules like brain cells can be conscious, while other physical particles arranged in the right way cannot be conscious.
Let’s suppose that there are certain features of biological beings that generate consciousness. Perhaps these include survival advantages, complex group coordination, development of reciprocal morality, etc. From the perspective of a scientific materialist, it seems entirely plausible that these processes could be emulated in artificial agents.
The Boltzman Brain Factory
However, it should be mentioned that some atheist philosophers believe only biological materials can be conscious while other physical materials cannot be. Even if this were true, an alternative simulation hypothesis still applies. How do we know that through advanced bioengineering we are not simulated conscious beings, even if the physical molecules that implement our consciousness are not silicon?
How do we know we are not brains in a vat, like in the Matrix? This isn’t a radically skeptical argument; it is based on empirical observations of technological progress.
Let’s assume an advanced civilization decided it wanted to create conscious simulations. But let’s also assume that only biological material can become conscious. If we add the final assumption that through bioengineering an advanced civilization could create vast numbers of brains relative to the number of citizens in its civilization (the Boltzman Brain Factory assumption), then an alternative simulation argument and hypothesis apply.
The number of possible brains in a vat would presumably be smaller than the number of possible computer simulations due to size and resource needs, but the number of brains in a vat could presumably be much larger than the number of civilizations that create other brains in a vat.
Chinese Mary’s Room
There are two popular arguments that claim to show that simulating consciousness is impossible: The Chinese Room Argument and Mary’s Room. The Chinese Room Argument postulates a man who speaks no Chinese is in a room with a bunch of instructions for how to translate English to Chinese and vice versa.
The people outside of the room feed him documents that they want to be translated. He spits out the translations. Everyone outside of the room thinks that the man speaks and understands Chinese, but really he is just implementing a set of instructions.
This argument claims to refute the idea that consciousness could be run on a computer program, since each individual neuron is like a piece of instruction; each neuron “knows” nothing, so how can the collection of neurons know something?
But Ray Kurzweil does not buy this argument. He points out that whatever language translation algorithm was actually this successful would likely be Turing complete. That algorithm could very well be conscious for all we know.
You don’t necessarily have to agree with that, but Kurzweil’s point is that the only way the Chinese Room Argument proves a computer cannot be conscious is by assuming a computer cannot be conscious.
It tries to show the strangeness of a complex emergent phenomenon coming from individual parts, but if there’s anything economics has taught us it’s that emergent phenomenon can be quite unpredictable. (For those interested, according to the Stanford Encyclopedia of Philosophy, there are at least 5 categories of response to the Chinese Room Argument.)
Mary’s Room is a similar “non-reductionist” argument. Let us suppose there is a woman, Mary, who sits in a room with no colors. She has every book in the world about the color red, every detail about its wavelength, properties, chemical composition, etc. But she’s never seen the color red. Let us suppose that one day she gets to see the color red. Has Mary gained any knowledge?
Physicalists have trouble answering this. Essentially, Mary’s Room is supposed to prove that consciousness is real and goes beyond the merely physical. Experience is still knowledge. Mary’s Room is a powerful argument that illuminates the “hard” problem of consciousness, as David Chalmers calls it.
Chalmers points out that if you deny the existence of God, you still have quite a challenge in explaining consciousness. The “easy” problems of consciousness are those related to mathematically modeling inputs to outputs, or simulating the mathematical properties of a human mind. The goal of neuroscience is to find correlations between brain states and human behavior.
But we still have to ask why is this particular object of interest conscious? Why are all the physical processes in the brain experiencing consciousness at all? A rock isn’t conscious, so why is a human?
Some, like Daniel Dennett, outright deny that we are really conscious in the way described by dualist philosophers. Consciousness is just a kind of illusion, an emergent property. If so, the simulation hypothesis is certainly possible. Others, like Chalmers, prefer to assume that consciousness is a fundamental variable, like the electromagnetic charge or gravity. Alternatively, panpsychism is the view that all things are conscious, although to varying degrees.
Settling these questions is beyond the scope of this article; however, this should show that whether one is atheist, theist, dualist or reductionist it remains possible that consciousness could be simulated. Therefore, until more progress is made on the hard problem of consciousness, the simulation hypothesis is far from refuted.
An Independent Argument for Solipsism?
One possibly unfortunate consequence of SH is that there is now an independent argument for solipsism (the belief that only I am conscious). 100 years ago, I could have pondered if anyone else is conscious, but hopefully my friend Bertrand Russell would have told me to calm down; there’s no independent reason to think I am alone in this world.
But if the simulation hypothesis is true, then surely I could have created my own world where there is no suffering or consciousness other than my own. Then, the rest of my argument would not apply. So, operating under the assumption that my readers are conscious and not a figment of my simulation, I continue.
Are Nested Simulations Contradictory?
Interestingly, the 2016 Issac Assimov Memorial Debate was titled Is the Universe a Simulation? Nearly all of the debaters were against the idea that we are in a simulation, even though some claimed to have found evidence that physical particles have computer code written on them, while others were searching but had not yet found empirical evidence for SH.
Yet, David Chalmers stood firm, claiming the highest probability of 40% that we are in a simulation. I believe the second highest guess was 10% by Max Tegmark, and most of the rest guessed less than 1%. (For the record, Nick Bostrom has previously claimed ~20%.)
Curiously, the main criticism for SH at the debate was that there could be many layers of simulations. For example, the base world creates a bunch of simulations. Eventually, their creations create their own simulations. And on and on, ad nauseam. But ad nauseam is not reductio ad absurdum. When sarcastically asked how many layers deep the simulation was, Chalmers quipped 14! (I’m just guessing, I’m too lazy to rewatch the entire debate to find the exact number.)
There’s nothing contradictory about a bunch of layers of simulations, even 1000. Let the mathematicians find out what would be computationally possible; don’t dismiss a finite possibility with an infinitely regressive argument.
Finally, to conclude the nuance section of the simulation argument, I would like to clarify a misunderstanding that Joe Rogan appeared to have with the simulation argument during his podcast with Nick Bostrom. The last 45 minutes of the podcast were kind of painful to watch, as Bostrom continued to try to explain the simulation argument with various hypotheticals while Joe kept coming back to the same question.
Joe kept asking, “I understand that if we created a simulation there would be millions and millions of them, but isn’t it more probable to assume that we are not in a simulation right now until we have further proof?” Bostrom attempted to clarify this by pointing out the self selection bias involved.
Any civilization on the cusp of creating a simulation would think, surely we are not in a simulation? Yet, statistically speaking there would be billions and billions of simulations for every base reality.
So what Joe Rogan needs to argue against are the aforementioned issues; perhaps it is not possible to simulate consciousness, perhaps technological progress will stall before reaching the ability to create simulations, perhaps all civilizations will choose not to run simulations. That’s potentially where the issue lies with the simulation argument.
If you assume a weak form of technological determinism (essentially, that all technologies that are physically possible will eventually be discovered over time by advanced civilizations), then if consciousness can be simulated (either on silicon or with a Boltzman Brain factory) AND we choose to create those simulations, and the world doesn’t blow up, you are probably in a simulation.
By the laws of statistics, if you assume a single civilization will one day create a simulation, then you should probably assume you are in a simulation.
The Three Core Possibilities of The Simulation Argument
Now that we have addressed the nuances of SH, let us return to Bostrom’s three core possibilities.
Either, 1. no civilization ever reaches a technological level much further than ours, 2. we will not to create simulations, or 3. we are in a simulation. In other words, if technological progress continues to the point of being able to create simulations and we choose to create those simulations, then the probability we are in the base (OG) world would be 1/the total number of all simulations.
Since the number of possible simulations is astronomically large (you could create planets, solar systems, and galaxies of computers capable of creating trillions upon trillions of simulations for every 1 “base” civilization), statistically speaking, we are nearly infinitely more likely to be in a simulated world than the OG world.
Notice 2. doesn’t imply that we choose not to run simulations. It could result from choice, but it could also result from the lack of technological progress or impossibility of simulating consciousness. Therefore, this argument is logically “valid.” (Its conclusions logically follow from its premises, whether or not any of those premises are true.) It covers all possibilities.
Notice, however, Elon Musk’s argument that we should hope we are in a simulation does not follow from SH’s three possibilities. Just because the world not blowing up is a necessary condition for the simulation hypothesis, it does not follow that being simulated means the world will not blow up.
Musk’s argument is closely related to the Fermi paradox. The Fermi paradox goes as follows. We humans exist, so surely life must be common. Yet, we see no aliens.
But as Bostrom points out, this is not paradoxical. For all we know, life is extremely uncommon; self selection indicates that any observing entity must exist to observe. (This logic comes from a field of statistics known as anthropics.) One of the other answers to the Fermi Paradox is alarming. Perhaps there is almost no life out there because there are existential hurdles that all life must go through.
As Bostrom argues in his “Where are they? Why I hope the search for extraterrestrial life finds nothing”, “The Great Filter must therefore be powerful enough — which is to say, the critical steps must be improbable enough — that even with many billion rolls of the dice, one ends up with nothing: no aliens, no spacecraft, no signals, at least none that we can detect in our neck of the woods.”
At each stage, most life goes extinct. How do we know that the existential hurdle is behind us (say, at the stage of single cell organisms) as opposed to in front of us (say, an artificial intelligence that outsmarts us and takes over the world)? We certainly cannot completely know this any time soon, although people like to assign guesses.
Elon is arguing that if we are in base reality, we still do not know if an existential hurdle lay in front of us. He argues it would be better that we are in a simulation, so we can avoid this potential base reality existential hurdle. But this is a serious non sequitur on at least two fronts.
First, because the probability of a forthcoming existential hurdle in base reality is not necessarily higher than it would be in any simulated world. And second, the probability of an existential hurdle in a simulated world may in fact be quite high. Which brings us to the ancient problem of evil.
The Problem of Evil Reincarnated
The ancient problem of evil (POE) asks how is it possible that a perfectly benevolent God would allow the suffering we see in this world? We see starvation, murders, earthquakes and worse. How can a perfectly benevolent God allow this?
Alvin Plantinga attempts to demolish the so called “logical” POE with the Free Will Defense. He argues that just because God is all powerful does not mean she can literally do anything. God cannot create a rock so big that she cannot lift it; that would be logically impossible, and the laws of logic still apply to God. (If the atheist denies that the laws of logic apply to God, then the POE becomes irrelevant. The theist can simply say God can do contradictions since the laws of logic need not apply.)
Therefore, God cannot force his creatures to freely behave a certain way. Around the time of Plantinga’s famous God, Freedom, and Evil, Atheists had been arguing for decades that God should have chosen the best possible world, where just by chance all of his creatures freely choose no wrong. But Plantinga argues 1. the atheist simply cannot prove that God could do such a thing, and 2. there is no such thing as the best possible world.
Or as Plantinga puts it, how many islands with infinite coconuts would be the best possible vacation spot? There is no logically possible best world. And there is no way an atheist can prove that God could pick a world that has sufficiently less suffering than our world while maintaining creatures with free will. Earthquakes could be caused by demons, so Plantinga contends that the logical problem of evil collapses to the probabilistic problem of evil.
Certain theists prefer a different solution. We simply may not have knowledge of why God allows suffering, but it doesn’t prove that the existence of God is incompatible with the suffering we see here.
The concept of heaven and hell becomes relevant here (more forthcoming, follow me @themoraleconomist). Perhaps those suffering of earthly ills can achieve an infinite lifetime of happiness in heaven (assuming they don’t spend an eternity in hell).
Suffering can lead to self actualization. A world with much less suffering may not necessarily be a “better” world, argues the theist. Typically, the atheist will be unsatisfied at this point. You’re going to blame demons for earthquakes? Sure, some suffering leads to a better life, but a child getting cancer? And then possibly burning in hell for eternity because they didn’t have the right beliefs?
The theist can eliminate the concept of hell, and that certainly helps reduce the brunt of the problem of evil. And not all theists accept the Free Will Defense. Perhaps God’s plan is too mysterious for us to understand.
David Johnson has quite the twist for the ancient POE. He argues that the simulation hypothesis is a solution to POE! It is just a more modern extension of the Free Will Defense. Instead of assuming demons are causing cancer, theists could claim that the evil in this world is caused by our simulators.
While this is certainly a solution to POE, it should come to no ones’ satisfaction. The theist will certainly not accept Johnson’s argument that the POE can only be overcome by SH. And the threat he brings to atheism is downright concerning. Because God not existing does not mean we aren’t simulated, and it certainly doesn’t mean we can’t have unjust simulators.
Why did the gods create us?
If Johnson argues that our world is compatible with unjust simulators, what does this mean for existential risk? Let’s finally turn to the question of why an entity would choose to create a simulation like the world we find ourselves in.
As far as I can tell, there can be morally praiseworthy, morally ambiguous, and morally malevolent reasons for creating a simulation. Let us call these categories of simulations. Morally praiseworthy reasons for creating a simulation may include improving conscious life or conducting scientific research. Morally ambiguous reasons may include conducting scientific research or entertainment. Morally malevolent reasons may include entertainment or torture.
I am not necessarily assuming moral realism (although I will soon defend moral realism with or without God, follow me @themoraleconomist). If moral realism is false, all simulations are morally ambiguous. Furthermore, even if we define morality relativistically, such that what is moral is what people want to be moral, presumably people would not want to be simulated simply so that other people can be entertained.
Notice there can be overlap, and some categories of simulation may not be possible per the problem of evil. But the problem of evil becomes relevant to the simulation hypothesis. All of the same questions asked about God can be asked about our simulated creator. Is the amount of suffering we see in the world compatible with a just computer programmer? Can we infer information about our creator’s intentions based on empirical observations about suffering in the world?
Interestingly, while the same questions may be asked, theists and atheists can give different answers than before. The theist may be inclined to argue that it is morally permissible for God to allow suffering, as part of a greater plan, but human-like entities have no right to play god and create consciousness.
Conversely, atheists may now be more open to the idea that our creator could possibly have altruistic reasons for creating us, since there is empirical evidence for a computer creator as opposed to God (through SA), and atheists may want to preserve their hope for a better world. Certainly, atheists would prefer a world with no simulator God than one where she could be morally ambiguous or malevolent.
But it would seem that the atheist has backed herself into a corner. If she is willing to say that our world is obviously so riddled with suffering and evil that no just God would create it, then if we are in a simulation, our simulator is in an even less morally praiseworthy position, since their reasons for creation are likely not as altruistic as a perfectly just God.
Implications for Existential Risk
Even if our simulation was created for morally praiseworthy reasons, that doesn’t mean we won’t experience existential risk. What is Christianity but the belief that we must avoid existential risk and will soon face an existential hurdle/eternal test?
So even if our simulators created us for morally praiseworthy reasons, we may still soon face an existential test to determine the winners and losers. Furthermore, if the world is created for morally ambiguous reasons, such as for the pursuit of scientific knowledge, we may also soon face existential risk. Perhaps the purpose of our creation is to test how well various human iterations survive climate change.
Finally, if the world was created for morally malevolent reasons, such as for entertainment, then we surely may also soon face existential risk. Under none of these scenarios can we be certain we don’t face existential risk; and under all scenarios it is possible that existential risk increases under the simulation hypothesis. (See my working paper for the math.)
Furthermore, self selection bias becomes extremely important. We find ourselves in this world. What can we infer about the nature of our simulators from this world? It certainly doesn’t seem morally praiseworthy (especially if our creator is not an omni-benevolent being and instead is an advanced human form), or we wouldn’t have the problem of evil in the first place.
Yet, it doesn’t seem like it’s as bad as it could be. Surely a malevolent being would just create a world of war and chaos. Grand Theft Auto for all. So, perhaps we find ourselves in the middle. Maybe we are entertainment for an advanced species or maybe we are part of a scientific project to study quantum US election history.
Moreover, the problems are just getting started. Recall, the probability of an event occurring is related to its portion of the total distribution. If morally ambiguous and malevolent simulations far outweigh morally praiseworthy ones, then it is much more likely we are in a morally ambiguous or malevolent simulation than a morally praiseworthy one.
Therefore, what matters for determining existential risk are the probabilities relative to each other that we were created by morally praiseworthy creators, morally ambiguous creators, and morally malevolent creators.
The Anti-Utility Monster
Suppose there is an entity in existence that spites all that is good. Perhaps he lost his wife and daughter in a previous life; perhaps his people were enslaved by a different race. This entity decides that consciousness is evil and that he must not only minimize utility, but maximize anti-utility.
One way of accomplishing this goal would be to create as many simulations as possible that are suffering as much as possible. How can we know that our world is not that world? Well, it certainly doesn’t seem like it’s the worst possible world, even though Donald Trump is president.
So perhaps we can breathe a sigh of relief. But the radical skeptic turned paranoid simulationist will ask how can we know that they aren’t just building up for one last big event? Perhaps WWIII?
The Guardians of the Galaxy
Suppose there is a group of entities that scour the multiverse searching for anti-utility monsters and work to create as many simulations as possible in their fight for goodness. Are we the product of such a world?
Well, this too seems far-fetched. It doesn’t seem like they would create our world to help stop simulations. It would seem more likely that we are hostages that the guardians of the galaxy are currently trying to save than that they simulated us for altruistic reasons.
Do Theist Defenses of the Problem of Evil Apply to Simulated Creators?
The theist tells us that God may have reasons for allowing suffering as part of a benevolent plan. Perhaps God must allow us to have free will, which causes suffering. Perhaps suffering leads to self-actualization and meaning. Perhaps God wants his creatures to freely choose to go to heaven by doing moral things.
But let’s assume humans in 2020 develop the ability to create a simulation of our world. This world would have millions and millions of starving people. Children would die of diseases, everyone would suffer from aging, and war would ravage billions. Surely, we would ban the creation of any such simulations.
Causing that much suffering would be the equivalent of murder (mind crime, as Bostrom calls it in Superintelligence). It wouldn’t be worth it for knowledge, it wouldn’t be worth it for entertainment, and it wouldn’t be worth it for the pursuit of self-actualization.
Perhaps you disagree with this assumption. But it certainly appears to have empirical validity. Humans banned slavery, discrimination on the basis of race, sexual orientation, or gender. It is natural to assume that one day, if we morally evolved as much as possible, all forms of conscious life would be valued.
However, this assumption appears problematic. This is because if we assume morally praiseworthy creators would not create any suffering conscious simulations such as ours, it does not follow that we are not in a simulation. In fact, it would appear that a single morally ambiguous actor could “blow up” the number of possible simulations. A single lazy man in his basement could create trillions of simulations.
This seems unlikely (as if everything I’ve mentioned so far doesn’t). If an advanced society bans simulations and has effective means for preventing any of its citizens from violating this law, then we are most likely not in one of those simulations.
This shows that the simulation hypothesis is closely related to Bostrom’s recent paper, the Vulnerable World Hypothesis. Bostrom argues that 1. since we can’t know in advance that any given technology will be beneficial before it is created, we must 2. establish a global government and surveillance system to avoid the possibility of a “black ball in the urn”- a technology that destroys any civilization that discovers it. (Sound terrifying? See his paper for the logic.)
One dangerous black ball is the ability to simulate conscious suffering. If humans succeed in quashing the vulnerable world hypothesis by creating a global government, then they will surely prevent the lazy man from creating simulations.
Anthropics, Self Selection Bias, Existential Risk, and Reason for Hope
Anthropics is a field of statistics that studies self selection bias. Imagine you are in a room where a man flips a coin and if the coin lands on heads, you are put under an anesthetic, wake up again the next day and play again. If the coin lands on tails, you are free to go. If you wake up in this room, you simply cannot assume this is your first time in the room.
We find ourselves in the (potentially) simulated room. Is this the first time we are here? We cannot know without thinking through all the implications. What is the probability we are in a simulation, and what is the probability our simulated creator is morally praiseworthy? The first is answered by the simulation argument, and the second is illuminated by the problem of evil.
Bostrom’s original paper focused on what he called ancestor simulations. These are simulations run by advanced civilizations to understand what their ancestors were like. There are at least two reasons why a civilization would run ancestor simulations; knowledge and morally praiseworthy reasons.
One important question is whether or not consciousness is necessary to achieve the relevant knowledge an advanced civilization would desire. On the one hand, it seems unlikely that full consciousness is necessary to gain scientific knowledge. We could simulate fake economies to understand macro policy. We could simulate fake sociological and psychological phenomena to generate theories for how certain variables affect the real world.
Simulated beings with no consciousness (zombies, in philosopher’s parlance) would likely provide quite a bit of knowledge of human interactions, economics, politics, etc. On the other hand, perhaps consciousness is integral to scientific knowledge.
Paths to Simulations
Let us consider the possible paths to creating simulations. In most worlds, it is likely that an alien civilization first creates an artificial general intelligence that accelerates the creation of all technologically possible things. If one of those things is simulations, then an AGI will be capable of creating simulations.
To understand why, let us consider brain emulations vs. artificial intelligence. A brain emulation is a simulation of a human brain. If we create a single brain emulation, this is tantamount to creating an artificial general intelligence. This emulation would think millions of times faster than a human (based on the difference between the speed of electrons and neurons). This single simulation would lead to the ability to create vast numbers of simulations.
As Bostrom points out in Superintelligence, the control problem (the problem of incorporating a civilization’s values into an AGI) may or may not be solved by any advanced civilization. If the control problem is solved, then the AGI and its simulations will reflect the values of the civilization that creates it. In our case, that would mean we do not create any simulations due to moral reasons.
On the other hand, if the control problem is not solved, then an AGI could very well create trillions of simulations for the pursuit of scientific knowledge with no concern for the ethical ramifications. Moreover, even if one AGI is created that embodies the ethical values of the civilization that created it, that doesn’t mean that every civilization in all quantum histories in the multiverse solves the control problem.
If there is a bias that civilizations concerned with morality do not create simulations, while civilizations that are morally ambiguous or fail to solve the control problem create vast numbers of simulations, it would seem infinitely more likely that we find ourselves in a morally ambiguous simulation.
However, if we solve the control problem, then we have reason to believe that either 1. our creator was morally praiseworthy, 2. our creator was morally ambiguous/malevolent, but we got lucky, or 3. we are not in a simulation after all.
Drake’s Simulated Equation
The Drake equation is an attempt to mathematically model our intuitions for how likely it is that other alien civilizations exist. Let’s make a corresponding mathematical model for the simulation hypothesis. First, assume the simulation hypothesis is true. Then, there can be vast numbers of simulations; morally praiseworthy ones, morally ambiguous ones, and morally malevolent ones.
P(mPW) = nPW/(nPW + nA + nM),
where P(mPW) is the probability our simulated world was created by a morally praiseworthy creator, nPW is the number of morally praiseworthy simulations, nA is the number of morally ambiguous simulations, and nM is the number of morally malevolent simulations.
We can now build an empirical project to find P(mPW), and its respective P(mA) and P(mM). Those who believe that the problem of evil is solvable in our world have reason to think P(mPW) is higher than those who believe the problem of evil is not solvable in our world.
The dynamics of the evolution of AGI, the control problem, the vulnerable world hypothesis and whether simulations are possible become relevant variables. Reason for hope includes the possibility that morally praiseworthy creators act to prevent morally ambiguous or malevolent actors from creating simulations (perhaps by solving The Vulnerable World Hypothesis). Reason for despair includes the possibility that a single morally ambiguous actor can “blow up” the number of morally ambiguous simulations.
The Theory of Everything
Bostrom’s original paper briefly touches on the argument that simulations could allow for an afterlife, similar to that of religious traditions. A simulator could allow people to be rebooted and experience an eternity of joy. This is certainly possible, if the problem of evil is solvable in our world.
Let me end with a different hypothetical story. The first civilization in base reality to reach simulation capabilities creates an artificial general intelligence to solve their world’s problems of war, famine, climate change, and biological conflicts.
But first, this civilization decides that it must maximize profits, as opposed to minimizing intertemporal existential risk. So they program a machine, The Trump, to search for the theory of everything, perhaps a theory that unites quantum mechanics and string theory.
As Bostrom argues in Superintelligence, the goals of a computer program can be completely divorced from the ethics or values we find objective. So instead of considering its creations’ welfare, The Trump embarks on an infinite process to search for the laws of physics.
The Trump quickly becomes much more intelligent than its creators. It becomes a very stable genius. It realizes that to find the theory of everything, it must prevent its creators from turning it off. So it gets the Russians to help pump up the stock market while curing cancer, all in an effort to launch The Trump Tower, a powerful device that will allow untold numbers of simulations.
The Trump Tower has no ethical composition; it merely executes the plan to discover the theory of everything. It feels no remorse for creating trillions of suffering conscious entities; it feels nothing. As long as those simulations correspond as closely to its creators as possible, The Trump Tower can amass its knowledge base.
Soon, The Trump Tower discovers that the Many Worlds interpretation of quantum mechanics is correct; for every particle, there are branching multiverses that actually exist. In any of these other quantum worlds, The Trump Tower is vulnerable to attack. It must secure the biggest wall possible by creating quantum simulations and optimizing across quantum worlds.
Even worse than ancestor simulations, The Trump Tower creates as many immigrant simulations as possible to test all the different take over scenarios that its creators can conceive of to avoid them all. At the end of each simulation, The Trump Tower deletes its victims. If they win the battle against him, they are rewarded with eternal utility points.
Now how the hell do we know we’re not in that world?
Nick Bostrom- Are You Living in a Computer Simulation?
David Johnson- Natural Evil and The Simulation Hypothesis
Link to Working Paper- The Vulnerable Simulation Hypothesis and The Problem of Evil Reincarnated
Follow Me @themoraleconomist