On Bostrom's Argument That You May Well Be a Computer Simulation

Nick Bostrom argues, in a 2003 article, that there's a substantial probability that we're living in a computer simulation. One way to characterize Bostrom's argument is as follows:

First, let's define a "post-human civilization" as a civilization with enough computing power to run, extremely cheaply, large-scale simulations containing beings with the same general types of cognitive powers and experiences that we have.

The argument, then, is this: If a non-trival percentage of civilizations at our current technological stage evolve into post-human civilizations, and if a non-trivial percentage of post-human civilizations have people with the interest and power to run simulations of beings like us (given that it's very cheap to do so), then most of the beings like us in the universe are simulated beings. Therefore, we ourselves are very likely simulated beings. We are basically Sims with very good AI.

Bostrom emphasizes that he doesn't accept the conclusion of this argument (that we are probably sims), but rather a three way disjunction: Either (1.) Only a trivial percentage of civilizations at our current technological stage evolve into post-human civilizations, or (2.) only a trivial percentage of post-human civilizations are interested in running simulations of beings like us, or (3.) we are probably living in a computer simulation. He considers each of these disjuncts about equally likely. (See for example his recent Philosophy Bites interview.)

Bostrom's argument seems a good example of disjunctive metaphysics and perhaps also a kind of crazyism. I applaud it. But let me mention three concerns:

(A.) It's not as straightforward as Bostrom makes it seem to conclude that we are likely living in a computer simulation from the fact (if it is a fact) that most beings like us are living in computer simulations (as Brian Weatherson, for example, argues). One way to get the conclusion about us from the putative fact about beings like us would be to argue that the epistemic situation of simulated and unsimulated beings is very similar, e.g., that unsimulated beings don't have good evidence that they are unsimulated beings, and then to argue that it's irrational to assign low probability to the possibility we are sims, given the epistemic similarity. Compare: Most people who have thought they were Napoleon were not Napoleon. Does it follow that Napoleon didn't know he was Napoleon? Presumably no, because the epistemic situation is not relevantly similar. A little closer to the mark, perhaps: It may be the case that 10% of the time when you think you are awake you are actually dreaming. Does it follow that you should only assign a 90% credence to being awake now? These cases aren't entirely parallel to the sims case, of course; they're only illustrative. Perhaps Bostrom is on firmer ground. My point is that it's tricky epistemic terrain which Bostrom glides too quickly across, especially given "externalism" in epistemology, which holds that there can be important epistemic differences between cases that from the inside seem identical. (See Bostrom's brief discussion of externalism in Section 3 of this essay.)

(B.) Bostrom substantially underplays the skeptical implications of his conclusion, I think. This is evident even in the title of his article, where he uses the first-person plural, "Are We Living in a Computer Simulation?". If I am living in a computer simulation, who is this "we"? Bostrom seems to assume that the normal case is that we would be simulated in groups, as enduring societies. But why assume that? Most simulations in contemporary AI research are simulations of a single being over a short run of time; and most "sims" operating today (presumably not conscious) are instantiated in games whose running time is measured in hours not years. If we get to disjunct (3), then it seems I might accept it as likely that I will be turned off at any moment, or that godzilla will suddenly appear in the town in which I am a minor figure, or that all my apparent memories and relationships are pre-installed or otherwise fake.

(C.) Bostrom calls the possibility that only a trivial portion of civilizations make it to post-human stage "DOOM", and he usually characterizes that possibility as involving civilization destroying itself. However, he also gives passing acknowledgement to the possibility that this so-called DOOM hypothesis could be realized merely by technologically stalling out, as it were. By calling the possibility DOOM and emphasizing the self-destructive versions of it, Bostrom seems to me illegitimately to reduce its credibility. After all, a post-human civilization in the defined sense, in which it would be incredibly cheap to run massive simulations of genuinely conscious beings, is an extremely grand achievement! Why not call the possibility that only a trivial percentage of civilizations achieve such technology PLATEAU? It doesn't seem unreasonable -- in fact, it seems fairly commonsensical (which doesn't mean correct) -- to suppose that we are in a computational boomtime right now and that the limits on cheap computation fall short of what is envisioned by transhumanists like Kurzweil. In the middle of the 20th century, science fiction futurists, projecting the then-booming increase in travel speeds indefinitely into the future, saw high-speed space travel as a 21st century commonplace; increases in computational power may similarly flatten in our future. Also: Bostrom accepts as a background assumption that consciousness can be realized in computing machines -- a view that has been challenged by Searle, for example -- and we could build the rejection of this possibility into DOOM/PLATEAU. If we define post-human civilization as civilization with computing power enough to run, extremely cheaply, vast computationally simulated conscious beings, then if Searle is right we will never get there.

Thoughts (B) and (C) aren't entirely independent: The more skeptically we read the simulation possibility, the less credence, it seems, we should give to our projections of the technological future (though whether this increases or decreases the likelihood of DOOM/PLATEAU is an open question). Also, the more we envision the simulation possibility as the simulation of single individuals for brief periods, the less computational power would be necessary to avoid DOOM or to rise above PLATEAU (thus decreasing the likelihood of DOOM/PLATEAU).

All that said: I am a cosmological skeptic and crazyist (at least I am today), and I would count simulationism among the crazy disjuncts that I can't dismiss. Maybe that's enough for me to count as a convinced reader of Bostrom.