Simulation Argument (Bostrom)

The thesis in one paragraph

Suppose a "post-human" civilisation has the computational power to simulate ancestor minds in arbitrary detail. Suppose even a small fraction of such civilisations choose to run such simulations, and that each runs many. Then simulated minds vastly outnumber pre-simulation biological minds. Apply the principle of indifference among indistinguishable observers: a randomly chosen mind that cannot tell whether it is biological or simulated should assign a high probability to being simulated. We can't tell. Therefore — assuming the empirical premises — we should assign a high probability to being simulated.

The argument's force depends on two conditional premises holding. Bostrom's actual conclusion is a disjunction: at least one of the following is true.

The trilemma

1. (Doom) Almost all civilisations at our level of development go extinct before reaching the technological capacity to run ancestor simulations.
2. (Disinterest) Almost all civilisations that do reach technological maturity choose not to run ancestor simulations.
3. (Simulation) We are almost certainly living in a computer simulation.

Bostrom's contribution is the structure: any one of these three may be true; the empirical work is calibrating which. He does not assert which holds. His later writings have leaned toward roughly equal probability across all three; others — Elon Musk, Neil deGrasse Tyson, David Chalmers — have publicly leaned toward (3).

The math

Let f_p be the fraction of civilisations that reach the post-human stage. Let f_I be the fraction of post-human civilisations interested in running ancestor simulations. Let N̄ be the average number of ancestor simulations run by an interested civilisation. Then the fraction of all human-like observers that are simulated is, approximately:

f_sim ≈ (f_p · f_I · N̄) / (f_p · f_I · N̄ + 1)

If the product f_p · f_I · N̄ is large (say, even 1000), then f_sim ≈ 1 — almost all observers like us are simulated. The argument's punch is that any non-negligible f_p · f_I · N̄ pushes f_sim close to 1, because computational resources on planetary or Dyson-sphere scales can support enormous numbers of simulations cheaply, swamping the original biological generation.

To deny (3) without denying the math, you must drive either f_p or f_I close to zero. Bostrom argues both look suspicious as near-zero values: many extinction risks exist but none is so universal as to wipe out essentially all civilisations; and "no post-human ever runs an ancestor simulation" requires extraordinary uniformity of values across vastly diverse posthuman cultures.

Positions in the debate

Position	Selects which disjunct	Author	Move
Doom selection	(1) Extinction	Many; Bostrom partly	Civilisations rarely survive to maturity
Disinterest selection	(2) Posthuman ethics	Bostrom (option)	Mature ethics forbids ancestor sims
Simulation realism	(3) Simulated	Chalmers 2022, Musk	Embrace; we are simulated
Computational impossibility	Premises false	Various physicists	Quantum simulation is intractable
Indifference rejection	Bayesian setup wrong	Weatherson 2003	Reject the principle of indifference
Reference-class objection	Setup wrong	Brian Eggleston	What counts as "observer like us"?
Russellian sim	Simulation isn't deception	Chalmers 2022	Even if simulated, our world is real

Worked example: a million Dyson swarms

Suppose by 3000 CE one civilisation has built a Dyson swarm around a single star, harvesting roughly 10²⁶ watts. Estimate that simulating one human-like brain at full subjective fidelity costs roughly 10¹⁷ operations per second (Kurzweil's number; others go higher). A Dyson swarm running at the Landauer limit could support around 10⁴² operations per second — enough for 10²⁵ simultaneously simulated minds. Even allowing many orders of magnitude in slack for safety margins and overhead, a single mature civilisation could run more simulated humans in a millisecond than have ever lived biologically (≈ 10¹¹).

The biological-to-simulated ratio thus inverts almost as soon as any civilisation reaches the relevant capability. The argument doesn't require many such civilisations — even one is enough to swing the population balance. To stop the conclusion, you need a reason why every advanced civilisation declines to do this, or no civilisation gets that far.

Counterarguments and modern responses

The simulation-shutdown lemma. Suppose we are in a simulation and we figure it out. Our simulators have plausible reasons to shut us down — to economise resources, to prevent us from probing simulation boundaries, or because we've stopped being interesting as an ancestor simulation. So if (3) is true and we publicly endorse it, our world may end. Robin Hanson's "How to Live in a Simulation" (2001) advises behaving in ways simulators are likely to find interesting. The lemma is not a refutation of Bostrom's argument but a strange-loop reason to think twice about acting on it.

Computational impossibility (Aaronson, Hossenfelder). Some physicists argue that simulating quantum mechanics at full fidelity scales exponentially with the number of particles entangled. To simulate even a small region of our universe with full quantum statistics would require a simulator larger than the simulated region. Bostrom's reply: simulators may compute on demand — render macroscopic detail at full fidelity, defer microscopic detail until it is observed. Whether this is empirically distinguishable from a non-simulation is the live question. Some physicists have looked for "lattice artifacts" in cosmic ray spectra; none found yet.

Indifference principle objections (Weatherson, others). The argument assumes a Bayesian observer with no evidence about which "type" of mind they are should assign probability proportional to the size of the reference class. But which reference class? "Minds like ours" is question-begging if "like ours" smuggles in features only biological minds have. The Doomsday Argument debates have shown that indifference reasoning over self-locating beliefs is fragile.

Disinterest from value drift. Bostrom himself canvasses the second disjunct: post-human civilisations may converge on values that forbid creating conscious sub-simulations. Suffering minimisation, computational ethics, simulated beings as moral patients — all could give post-humans reason to abstain. Critics object that even one renegade civilisation undoes the abstention; the argument requires universal disinterest, an extraordinarily strong prediction.

Chalmers's "simulation isn't fake" reply (2022 Reality+). Even if we are simulated, that doesn't make our world unreal. Chalmers argues that simulation realism is a structural hypothesis about what our world is made of — bits running on a computer, rather than fundamental physics — but the world itself, the trees, our friends, our minds, are real. The simulation argument isn't a skeptical conclusion that nothing matters; it's a metaphysical conclusion about what underlies what.

Reference-class manipulation. The argument compares simulated minds to biological minds. But what about Boltzmann brains, animal minds, AGI minds, hybrid minds? If the reference class of "minds like mine" is read narrowly, the calculation changes. Brian Eggleston and others have argued the indeterminacy of reference classes drains the argument of its punch.

Variants and descendants

• Tegmark's mathematical universe. Max Tegmark (2014, Our Mathematical Universe) takes the argument further: every mathematical structure that supports observers is actually instantiated. We are not in a particular simulation — we are a mathematical pattern.
• Chalmers's structuralism (2022). "Reality+" recasts the simulation hypothesis as a structuralist hypothesis: simulators have built our physics; the physics is real even if its substrate is silicon or its successor.
• Boltzmann brains. Statistical-mechanics version: in an infinite cosmos, random thermal fluctuations occasionally produce self-aware brains with arbitrary memories. By measure, such brains may outnumber evolved minds — pressing similar reference-class issues.
• Doomsday Argument (Carter, Leslie). Sister anthropic reasoning: if you assume you're a randomly chosen human in human history, and you find yourself relatively early or late, that updates your estimate of total human lifespan.
• Nested simulations. If we are simulated, our simulators may also be simulated. Stacking gives a tower with a real root somewhere — but you cannot tell which level you occupy.
• Selectively rendered worlds. A practical refinement: simulators conserve compute by rendering macroscopic detail and only deferring quantum detail when measured. Predicts subtle anomalies in unobserved domains.

Common confusions

• Bostrom does not say we are simulated. The published conclusion is a disjunction. Reading the paper as "Bostrom thinks we're in a simulation" misrepresents both the paper and his subsequent statements.
• The argument is not proof — it is conditional probability. If you reject the empirical premises (technological maturity reachable, ancestor sims sometimes run), the conclusion doesn't follow. The premises are calibrations, not theorems.
• "It changes nothing" is not a reply. Even if the simulation hypothesis has no day-to-day implications, that does not falsify it. Pragmatic indifference and metaphysical truth are independent.
• The simulation argument is not religion. It does not posit a benevolent or omniscient designer — only a sufficiently competent computer scientist. The simulators may be incompetent, indifferent, or running their simulation as a Tuesday lab assignment.
• Falsifiability is unclear. Standard science prefers falsifiable claims. Bostrom's argument is not strictly falsifiable from inside the simulation; some philosophers see this as a cost, others as the nature of metaphysical theses.