Are we in a simulation?

Almost everything written about the simulation argument gets its structure wrong. Nick Bostrom's 2003 paper does not argue that we are probably in a simulation. It argues that at least one of three propositions is true, and it leaves the choice between them open. Read as a prediction it is easy to dismiss; read as the disjunction it actually is, it is much harder to get out of.

What the argument actually says

The three propositions are these. Almost every civilisation at our stage of development goes extinct before reaching technological maturity. Or: almost every mature civilisation loses interest in running detailed simulations of its own ancestors. Or: we are almost certainly living in one. Bostrom's claim is that you cannot consistently deny all three.

The engine is a counting argument. Suppose even a small fraction of mature civilisations run ancestor simulations, and that a mature civilisation has the computing power to run very many of them — Bostrom's estimates put the cost of simulating a human mind at a rounding error against the resources of a planetary-scale computer. Then, among all beings with experiences like ours, the simulated ones vastly outnumber the original biological ones. Add an indifference principle — that you should reason as if you were a random member of the class of beings with your experiences — and you should conclude you are probably one of the many, not one of the few.

Why this belongs on a site about the anthropic principle

That last move is not a computing claim. It is observer selection, in exactly the form the rest of this site is about. The argument reasons from what kind of observer you are likely to be, given a population of possible observers — the same step the Doomsday Argument makes with birth ranks, and the same step that separates self-sampling from self-indication. Its strength and its weakness both come from there.

It inherits the reference-class problem along with the method. 'Beings with experiences like mine' has to be drawn tightly enough to include the simulated copies and loosely enough that you do not already know which side you are on. It also inherits the substrate assumption: the counting only works if a sufficiently detailed computation of a mind is a mind. That is a widely held view in philosophy of mind, but it is a position rather than a result, and someone who rejects it is not being inconsistent — they are denying a premise.

The objections that bite

Brian Weatherson's reply is the cleanest of the early ones. He argues the indifference principle Bostrom needs is stronger than the one he defends, and that under a weaker and more plausible version the conclusion does not follow: knowing that simulated observers outnumber real ones does not by itself tell you how to distribute your credence, because you have information about your own situation that the counting throws away.

David Kipping's 2020 paper does the arithmetic properly as a Bayesian model comparison, and it finds two things worth knowing. Under a flat prior the physical hypothesis comes out very slightly ahead — the odds are close to even, not the lopsided victory the popular version claims. And the balance is fragile in a specific direction: if our own civilisation ever produces a convincing ancestor simulation, the second proposition is falsified by demonstration and the posterior swings hard towards the third. That is the closest thing this subject has to a decision procedure.

Then there is the question of whether any version of the claim is testable. Silas Beane and colleagues looked for the signature a particular implementation would leave — a universe computed on a spacetime lattice should impose a small anisotropy and a cut-off on the highest-energy cosmic rays — and derived constraints. It is honest work, and its limits are the point: it tests one specific architecture, chosen because it is the one we know how to look for. A simulation that renders only what is observed, or that runs on principles we have not thought of, leaves no such fingerprint. The general claim is not falsifiable, and a hypothesis that survives every possible observation is not paying rent.

Older than the computer

The sceptical shape of this is ancient. Zhuangzi could not tell whether he was a man who had dreamt he was a butterfly or a butterfly dreaming he was a man. Descartes, in 1641, asked what he could still know if a malicious demon were feeding him a convincing false world, and answered: almost nothing, except that something was being deceived. Hilary Putnam gave the modern version its name with brains in a vat, and then argued it away — on his causal theory of reference, a brain that has only ever been fed simulated trees cannot successfully refer to real trees, so its sentence 'I am a brain in a vat' is false whenever it is uttered.

What Bostrom changed is the type of claim. The classical versions are sceptical hypotheses: unfalsifiable by construction, raised to test what knowledge requires, and never seriously believed by the people raising them. His is a statistical argument about a population, built out of premises about computing hardware and the future of technological civilisations, and it concludes with a probability rather than a doubt. That is what makes it an anthropic argument rather than an epistemological one — and it is also what exposes it to the objections above, which the demon was always immune to.

Its place among the four answers

In the fine-tuning debate the simulation argument shows up as the modern descendant of the design argument: a universe whose constants sit in a narrow habitable band because somebody chose them, with the designer relocated from theology to a computer lab. The structure is the same and so are the problems — it explains the data almost too easily, and it raises the question of what explains the simulators' universe, which has to be fine-tuned enough to contain them.

It differs from the design argument in one respect worth crediting: it has a mechanism that does not require anything supernatural, and one of its branches makes a concrete, if distant, empirical prediction about our own future behaviour. It differs from the multiverse in that it does not need any new physics, only a lot of computing. Whether that makes it a better answer or merely a more familiar-sounding one is exactly the kind of judgement the four answers are built to make explicit.

The honest verdict

The trilemma is valid. Its premises are contestable, and the two that matter — substrate independence and the indifference principle — are philosophical commitments rather than findings. If you accept both, you owe an answer to which of the three doors you are taking, and 'I do not know' is a respectable one: no observation currently distinguishes them, and it is not clear what would.

What the argument really demonstrates, and this is its lasting contribution, is how quickly reasoning that counts observers takes over once population size is allowed to carry evidential weight. That is the same machinery that produces the Doomsday Argument and the Presumptuous Philosopher, and the fact that it can deliver a conclusion this large from premises this small should make you look hard at the machinery — which is what the rest of this library is for.