Of all the arguments in this course, this is the one most people have heard and the one most people have heard wrong. The popular version is "we are probably living in a simulation, and a philosopher proved it". That is not the conclusion, it is not close to the conclusion, and the gap between the two is a useful lesson in reading an argument for what it says.
What the argument actually claims
Nick Bostrom published "Are You Living in a Computer Simulation?" in Philosophical Quarterly in 2003. Its conclusion is that at least one of the following three propositions is true:
- The human species is very likely to go extinct before reaching a posthuman stage of technological maturity.
- Any posthuman civilisation is extremely unlikely to run a significant number of simulations of its evolutionary history.
- We are almost certainly living in a computer simulation.
That is a disjunction. It says one of these holds, and it does not say which. Bostrom's own stated view has been that we should distribute our belief roughly evenly across the three, which is the opposite of a claim that the third is true.
The reasoning behind it is short. Suppose a civilisation reaches the point where it can run detailed simulations of conscious beings at low cost. Suppose some such civilisations choose to run simulations of their own ancestors, which is the kind of thing curious and wealthy societies do. Then the number of simulated observers with experiences like ours vastly exceeds the number of original ones, because the originals are one civilisation and the simulations are many. If you have no way of telling which kind you are, you should think you are probably of the more numerous kind.
Each of the three propositions is a way of blocking that conclusion. The first blocks it by denying anyone gets the technology. The second blocks it by denying anyone uses it. The third accepts it.
Example. A friend says: "Bostrom calculated the probability we are simulated at over 99 percent." What has gone wrong?
Two things. The argument produces no such number and produces no unconditional probability at all. What it produces is a conditional: if the first two propositions are false, then the third follows, because the counting argument goes through. Turning that into an unconditional figure requires you to supply your own confidence that civilisations survive and that survivors simulate, and neither of those is something Bostrom claims to know. The friend has taken the last horn of a trilemma and reported it as the conclusion of the whole thing, which is the standard misreading. Note the general shape, since it comes back in lesson ten: a conditional result presented as a categorical one is the commonest way a careful argument gets inflated in transmission.
Now you. Which of the three propositions does a committed techno-optimist who also thinks simulating conscious beings is morally monstrous end up accepting?
Answer
The second. They deny the first, since they expect civilisations to reach technological maturity, and they want to deny the third. That leaves them holding that mature civilisations, although able to run ancestor simulations, almost never do. Their moral objection supplies exactly the reason: if simulating conscious beings is a serious wrong, a mature civilisation might well prohibit it, and the prohibition would have to be near universal and near permanent for proposition two to hold, since it takes only a few defectors running many simulations to restore the counting argument. Seeing that the position is available, and seeing how demanding it is, is the point of the exercise.
The assumption that carries the weight
The argument needs one substantial premise that is easy to miss: substrate independence, the thesis that consciousness depends on the structure of a computation rather than on what the computation runs on. If a mind is what a certain pattern of information processing does, then implementing that pattern in silicon produces a mind, and simulated people are people.
If substrate independence is false, the argument collapses at once. A simulation would then contain no observers, only very good models of observers, and there would be nobody inside it wondering whether they were simulated. The counting argument needs simulated minds to count.
This is why lesson eight, on consciousness, is not a detour. The simulation argument is a hostage of the hard problem: it is only as strong as the claim that running the right program is sufficient for experience. Bostrom is explicit that he assumes this, and the assumption is not a small one. A large minority of philosophers of mind reject it.
There is also a quieter assumption in the counting step. It says that if most observers with your evidence are simulated, you should believe you are probably simulated. That is an indifference principle, and it is the same kind of reasoning that lesson thirteen examines under the name of anthropic reasoning, where it produces the Doomsday argument and a good deal of trouble. Whether the principle is sound is genuinely open, and anyone who accepts it here should be prepared to accept its consequences there.
There is a third assumption, less often noticed, about cost. The counting argument needs simulations to be cheap relative to the resources of a mature civilisation, since a civilisation that could afford three ancestor simulations would not swamp the original population. Bostrom estimates the computational requirements and argues they are modest for a civilisation with planetary-scale computing. The estimate depends on a further claim, that the simulation need not be run at full physical fidelity: it can model minds in detail and fill in the rest of the universe on demand, rendering the distant galaxy only when somebody points a telescope at it. That is a plausible engineering shortcut and it introduces a vulnerability, since a world built that way might behave oddly under sufficiently careful observation, which is what the testing proposals below are looking for.
Example. Suppose it turns out that consciousness requires specific biochemistry and cannot be implemented in silicon at all. Which of the three propositions does that establish?
None of them, and that is the point worth taking away. Substrate independence is not one of the three horns, it is a premise of the argument that generates them. If it fails, the trilemma never gets stated: there is no counting argument, because ancestor simulations contain no observers to count, and the question of which proposition is true does not arise. A reader who rejects substrate independence is not choosing horn one or horn two, they are declining the whole argument. Confusing a premise with a disjunct is easy here because horn two also concerns simulations not happening, but the two positions are quite different: horn two says nobody runs them, while rejecting substrate independence says running them would not produce anybody.
Now you. A defender replies that even if biological consciousness were required, a sufficiently advanced civilisation could grow biological brains in vats and feed them a simulated world, so the argument survives. Does it?
Answer
It survives in a weakened form and loses most of its force. The counting argument needs simulated observers to be enormously more numerous than original ones, and that multiplier came from simulations being cheap. Growing and maintaining physical brains is not cheap: it consumes matter, energy and space at something like the rate the originals did, so a civilisation might run a few such worlds rather than billions. With the ratio near one instead of astronomically large, the conclusion that you are probably simulated does not follow. This is a good illustration of how a premise can be technically rescued while the argument that rested on it quietly dies, and it shows that substrate independence was never doing merely decorative work.
What would count as evidence
The scientific press periodically reports that physicists have found a way to test the hypothesis. It is worth being clear about what such tests could and could not do.
Proposals usually look for artefacts of discretisation: a smallest length, a cutoff in the spectrum of high energy cosmic rays, a lattice structure that would betray a simulation running on a grid. One much-discussed paper by Beane, Davoudi and Savage (2012) worked out what signatures a lattice quantum chromodynamics simulation of our universe would leave.
The trouble is asymmetric. A positive result would be striking but not decisive, since a discrete universe is a live physical hypothesis in its own right and needs no simulator. A negative result shows only that we are not in a simulation of that particular design, and the space of designs is unbounded. A simulator with any competence would not leave detectable seams, and a simulator with the resources to run our world can presumably patch the parts we look at closely.
That last consideration is the interesting one, and lesson nine takes it up properly. A hypothesis that can absorb any observation whatever is in trouble, not because it is false but because its flexibility means no observation supports it either. The simulation hypothesis is unusually good at absorbing observations, and its defenders often treat that as a strength. It is a weakness, and saying exactly why is the business of lesson nine.
Example. Someone argues: "The universe is made of quantised fields, spacetime may be discrete at the Planck scale, and quantum mechanics has an observer-dependent flavour. That is what a simulation would look like." What is wrong with the inference?
It treats features that are equally well explained by physics being as it is as though they were predictions of the simulation hypothesis specifically. To count as evidence, an observation must be more likely if the hypothesis is true than if it is false, which is the machinery of lesson five. Quantisation is exactly what you would expect from a quantum field theory whether or not anyone is simulating it, so it does not discriminate. The observer-dependence point is worse, since it rests on a popular misreading of measurement in quantum mechanics that does not require anything mental. Notice that the argument would have been made in the same tone had the universe turned out continuous, with continuity cited as evidence of unlimited computational resources.
Now you. Suppose the argument's counting step is granted, along with substrate independence. Does it follow that the world around you is unreal?
Answer
No, and the confusion is worth naming because it drives most popular discussion. If you are in a simulation, your chair is a pattern in a computation rather than a pattern in a quantum field, but it is still a real chair in the only sense you ever cared about: it reliably holds you up, other people bump into it, and it persists when you leave the room. What changes is the story about what the world is made of, not whether there is a world. This is the same distinction lesson seven draws between idealism and scepticism about the external world, and it is why the simulation argument, despite the vat comparison, is not really a sceptical argument at all. It is a hypothesis about the ultimate substrate, which is a metaphysical claim, and it leaves your ordinary knowledge exactly where it was.
Why it belongs in this course
Three reasons, none of which is that the conclusion is likely.
First, it is the clearest available case of a valid argument with a surprising conclusion, where the work of responding is choosing a premise. Everyone who has an opinion about it has, whether or not they realise it, accepted one of the three propositions.
Second, it shows how much a single hidden assumption can carry. Substrate independence is stated in one sentence in Bostrom's paper and everything depends on it.
Third, it is the best illustration in the course of the difference between a claim you cannot rule out and a claim you have reason to believe. You cannot rule this one out. That fact alone gives it nothing, and the reason it gives it nothing is the subject of lesson five.
The sceptical arguments are now on the table in their classical and modern forms, stated at full strength and unanswered. The next lesson answers them, or rather surveys the four serious attempts and prices each one.