Refutations of the Simulation Argument

Fabien Besnard, [email protected]

5th September 2004

Abstract
By examining the logical consistency of Nick Bostrom’s simulation
argument, we find that its conclusions are not fully compatible with
its premises, a weak form of the liar’s paradox. We also claim that a
flaw in the simulation argument is to be found in the misuse of finite
probability theory.

1 Introduction
The argument developed in [1], known as the simulation argument, has re-
cently attracted much attention. Even some popular science magazines like
[3] have accounted for it. According to this argument, a three-way disjonc-
tion must be true. More precisely, at least one of the following statements
must be true : (1) almost no civilization of human level ever reach a stage
of very advanced technology, called post-humanity and characterized by the
ability to run a computer simulation of a whole human-level civilization,
(2) almost all civilizations that reach post-humanity do not desire to run
computer simulation of human-level civilizations, or (3) we probably live in
a computer simulation.
Since the consequences of this argument are rather dramatic, it must
be investigated in details. Authors have already criticized the argument
on the ground on its rather wild implications, perhaps containing the seeds
of its own invalidity. P. Davies even sees it as a reductio ad absurdam of
multiverse theories ([4]). However, it must be noted that the conclusion
of the simulation argument is not (3), but only a three-way disjunction
involving (3). Is it enough to allow one to discard the argument in the name
of rationality ? At any rate, since the simulation argument is claimed to be
a rational argument, it is important to inspect its logical and mathematical
aspects. We propose to tackle to this task in this article.

1
 The simulation argument relies on several assumptions such as the sub-
strate independence and the indifference principle that we do not wish to
discuss and take for granted in the following. For convenience we summarize
the simulation argument in the second section. Then, in the third section
we use a strategy inspired by the liar’s paradox to reduce the argument to
the absurd. We conclude that there must be a flaw in the argument, and
give two related propositions about it in the fourth section. We call them
the cardinality problem and the viewpoint problem.

2 Summary of the simulation argument

The simulation argument consists in two parts. In the first, which we will
call the foundation, the author, relying on several technical conjectures in
computer science, claims that it is highly probable that an enormous quan-
tity of computing power should be available to sufficiently advanced civi-
lizations. These advanced civilizations are called ’post-human’. It is argued
that post-human civilizations, or individuals in them, could use the immense
computing power they possess to simulate human-level civilizations like ours,
though only a fraction fI of them would be interested in doing that. We call
NI the average number of simulations that each interested post-human civ-
ilization would then creates. Because of the enormous computing power of
post-human civilization, NI is huge. This is the conclusion of the foundation
part of the argument.
In the second part, called ’the core’, the author computes the fraction of
all human-level civilizations that are being simulated. This is :

fP fI NI
fsim =
(fP fI NI ) + 1
where fP is the fraction of human-level civilization that reach post-humanity.
Then, it is easily shown that at least one of the following is true :

fp ≈ 0 (1)

fI ≈ 0 (2)
fsim ≈ 1 (3)
The interpretation of (3) is that we probably live in a computer simulation.
The simulation argument can thus be summarized by the following dia-
gram :
F −→ NI >> 0 =⇒ (1) or (2) or (3) (∗)

2
The foundation is ’F −→ NI >> 0’1 , and the core is ’NI =⇒ (1) or (2) or (3)’.
For our purpose, all we need to know is that F depends on the currently
known laws of physics and some conjecture about the technical feasibility of
very complex computer simulations.

3 Incompatibility of (3) with the foundation of the

argument
Let us suppose that some piece of evidence (S) has been found that make us
suspect that (1) and (2) are false, while (3) is true. Then let us look at the
line of reasoning we get if we throw (S) in the hypotheses, living all other
parts of the simulation argument unchanged :

F and (S) −→ NI >> 0 =⇒ (1) or (2) or (3) (∗∗)

Logically if some theory makes a prediction, and this prediction is real-
ized, at the very least the confidence we have in the theory cannot lessen.
But in fact (**) is a very doubtful argumentation. Indeed, given (S), the
confidence we have in the laws of physics would be much lessened. In his
article Bostrom emphasizes that the laws of physics may be very different
than the ones we know of, if we happen to be simulated. Also, it is very
possible that there is an automatized procedure that detects any human-
made experience which would reveal an intrinsic granulosity of space or
time ([1]). Actually such a granulosity would breaks Lorentz invariance,
because a minimal length cannot be Lorentz-contracted. Recent theories of
quantum gravity actually predict such a phenomenon, and upper bounds on
a minimal unit of lenght have already been deduced from experiment ([2]).
This means, according to Bostrom, that the results of these experiments
could very well have been forged by ’those who simulate us’. It is easy to
see that in such conditions, a radical form of scepticism should prevail. In
particular, predictions about what could be achieved in the future would not
be very reliable (if ever such predictions can be). Actually, even our belief
that the sun will rise tomorrow would be diminished, since the simulation
could be terminated at any moment. We claim that in view of this, the
credence we can have in (**) must be necessarily less than the credence we
can have in (*). Thus, (*) would turn out to be a theory that becomes less
and less credible as its predictions are realized ! It does not mean that in
1
We recall that >> 0 means ’very big’

3
such a situation we should not believe to be in a simulation, but that this
belief would only be based on (S) and not at all on the simulation argument.
In view of the preceeding paragraph, it appears that the simulation ar-
gument cannot be correct if (1) and (2) are false. It means that among the
eight possible triples of truth-values (1), (2) and (3) can take according to
the conclusion of the simulation argument, at least one (false, false, true)
is incompatible with the foundation part of this argument. This does not
imply that the simulation argument is false, but at least it is incomplete.
Let us look for example at the following deduction :

x2 = 1 and x > 0 =⇒ x = 1 or x = −1

It is perfectly true from a logical point of view, but obviously something has
been forgotten because one part of the final disjonction is incompatible with
one of the premise. We would say that this deduction is incomplete, and
even misleading because it forces us to consider a case that cannot happen.
Of course, once the incompatibility has been found, the deduction is easily
refined by removing x = −1 from the conclusion.
By carefully inspecting the simulation argument, it appears that every
step of the argumentation seems compatible with the hypothesis of (1) and
(2) being false. Perhaps this is not so and we are not clever enough to see
it. In this case, the simulation argument could be refined, thus leading to
conclusion that (1) or (2) must be true. It means that the ’simulation part’
of the simulation argument would have to be removed !
In any case, there must be a flaw somewhere in the simulation argument.

4 Possible flaws in the core of the argument

4.1 The cardinality problem
A fraction is the quotient of two numbers. To calculate fsim , fI and fP ,
we need the cardinal of the set of all post-human civilizations, all human-
level civilizations, all simulated human-level civilizations, all post-human
civilizations interested in simulating human-level civilizations. The question
that naturally arises is ’are these sets finite ?’. The biggest set that appears
in the argument is the set of all civilizations whether human or post-human,
simulated or not. In probability theory this set would be called the universe,
but here it would be more appropriately called the multiverse. Let us call it
Ω. It is not needed to know the precise structure of Ω as long as it is a finite
set, because one just has to assume equiprobability of every element of Ω to

4
do the calculation. On the other hand, if Ω might be infinite we would have
to prescribe a certain probability law2 on it, thus its structure would have
to be known. We see that it is crucial for the argument that Ω be a finite
set. In the following we will see that in some cases, it is not so.
At the end of his paper, N. Bostrom consider the possibility that a sim-
ulation could take place inside another simulation. Actually, this possibility
must be considered, because it is a consequence of a recursive application of
the simulation argument. Therefore the multiverse could have a ’forest-like’
structure, like in the following figure.

...

...

A possible structure for the multiverse

The root of each tree represent a ’real’ post-human civilization, and

each node represents a simulated civilization, which can be post-human and
simulate other civilizations as well. The root without branches represents a
human non-simulated civilization. As N. Bostrom puts it, we cannot know
for sure what the laws of physics at each node or root look like. If they are
very different from the ones we know of, the possibility that a node may
generate an infinity of branches cannot be ruled out, neither the possibility
that there may be an infinite number of trees in the ’forest’. In either case
Ω is infinite and the argument is flawed.
Suppose on the contrary that there are only a finite number of trees, let
us say just one for the sake of simplicity. Then, for the tree to be finite,
it is necessary that each node of the tree gives birth to a finite number
of branches only and that there is no path of infinite length in the tree3 .
It could appear reassuring because if the root has only a finite amount of
computing power, an assumption that sounds reasonable, it seems at first
sight that it cannot generates an infinity of branches at the first level, neither
2
or at least a measure.
3
this is König’s lemma of graph theory.

5
sustain an infinitely high tower of simulations. Now let us ask the question :
’what time is it ?’. It is a very important question because the definition of Ω
does not refer to any notion of time. Thus, the root could run a simulation
s1 , from its begining to its end, then a simulation s2 and so on, and the
resulting tree (representing all the simulations ever generated by the root)
would have an infinity of branches. It would even be possible for the root
to simulate an infinity of infinitely long lasting simulations, just by running
s1 for a unit of time, then s1 and s2 each for a unit of time, then s1 , s2 and
s3 again for one time unit, and so on. The simulations would run more and
more slowly but nothing could be suspected ’from within’. We do not find
any reason to exclude the possibility that a post-human civilization live for
ever. In this case, Ω is again infinite. In the end we find that having an
infinite computing power or living for ever amounts roughly to the same for
a post-human civilization.
While the above does not prove that Ω is infinite, it casts serious doubt on
the possibility of doing the simple calculations of the ’core’ of the simulation
argument 4 .

4.2 The viewpoint problem

There is an even worse problem in the definition of Ω : it is not obvious that
it is a set ! To answer this question, we must wonder what is our point of
view when we speak of all civilizations. This collection cannot of course be
gathered before our eyes, even in principle. Therefore we cannot describe it
as a part of the physical world. The only rational tool that remains is then
pure mathematics. But we know since B. Russell that describing something
by its properties is not a valid way of defining a set. For instance, the set of
all sets does not exists, neither does the set of all sets that have some fixed
mathematical structure (like the set of all vector spaces). The usual way of
defining a set is to describe it as a subset of some known set. Since we lack
a mathematical description of the multiverse, we cannot affirm that it is a
set.
We think that the problem of viewpoint is the key to understand the
flaw in the simulation argument. The latter begins with rather plausible
assumptions about the world we know and love, which we follow without
noticing a rather subtle change in the point of view, and at the end we have
gotten completely off the ground. We think that for an argument aiming
at calculating the probability that we live in some part P of a bigger set Ω,
4
It should also be noted that not only Ω but several subsets of Ω appearing in the
calculation may be infinite, and NI as well in some cases.

6
the definition of Ω should be made clear from a mathematical point of view.
This is a difficulty that cannot be avoided : it is well known in probability
theory that the set of all possible outcomes must be prescribed before any
calculation is undertaken. Once this problem has been raised, we see the
rather wild implications of the mere existence of Ω and this, in our mind,
forbids the argument to go any further. More simply put, if we imagine
from the begining that we may be in a simulation, the argument cannot go
further. At least it must be ammended.

5 Conclusion
The method we have used in section 3 makes the simulation argument appear
very reminiscent of the liar’s paradox, though in a weaker form that does not
give rise to a full-fledged paradox. A statement like ’I can prove that nothing
can be proved’ is easily seen to be incompatible with its own conclusion and
our method would apply to it. Like the simulation argument, it can be
given the truth-value ’false’. However the simulation argument is not an
argument of formal binary logic. As such, it need not have a truth-value.
However it should be noted that in our method we have only made use of
the reliability one could assign to the simulation argument, and shown that
it would decrease if some of its predictions turned out to be realized.
One might be tempted to conclude likewise that any so-called rational
argument whose conclusion would be incompatible with rational thinking is
to be rejected. However, it should be noted that we have not claimed nor
used such a strong principle in our refutation of the simulation argument.
We have carefully split the argument in two parts : the foundation part
that is based on physics and more generally on the natural sciences, and the
core, which is based on mathematics. We have argued that the validity of the
conclusion of the simulation argument is not compatible with its foundation
part, but we have not claimed that it spoils any form of rational thinking,
including mathematics, since it could be argued that it does not (a question
that sounds like ’Does God have the power to make Pythagoras’ theorem
false ?’).
It should also be noted that our criticism about the mathematical part
of the simulation argument is logically independent from our rational ques-
tioning of section 3. Thus we believe to have provided two independent
refutations of the simulation argument. However, it is very interesting to
note that the inspection of the mathematical foundation of the argument re-
veal a ’viewpoint problem’ that also looks like the ’liar’s problem’ of section

7
3.
Of course, the physical part of the argument by itself is certainly not
immune to critcism. We believe that even if a simulation of the physical laws
is conceivable to a certain extent, the complexity of the initial conditions for
a good simulation is a problem that cannot be overestimated. However, we
do not feel competent to argue on this point, and leave it to others.

6 References

References
[1] N. Bostrom, ”Are you living in a computer simulation ?”, Philosophical
Quarterly vol. 53 (2003), No 211, p. 243-255.

[2] G. Amelino-Camelina, ”Are we at the dawn of quantum-gravity phe-

nomenology ?”, Towards Quantum Gravity, Lecture Notes in Physics,
541, p.1-44

[3] J.P. Delahaye, ”Sommes-nous réels ?”, Pour la Science, August 2004.

[4] P. Davies, ”Fitness and the cosmic environment”, in Fitness of the cos-
mos for life : Biochemistry and fine-tuning, too appear.