How to Live in a Simulation
Abstract
Those who believe suitably programmed computers could enjoy conscious
experience of the sort we enjoy must accept the possibility that their own
experience is being generated as part of a computerized simulation. It would be a
mistake to dismiss this is just one more radical sceptical possibility: for as
Bostrom has recently noted, if advances in computer technology were to continue
at close to present rates, there would be a strong probability that we are each
living in a computer simulation. The first part of this paper is devoted to
broadening the scope of the argument: even if computers cannot sustain
consciousness (as many dualists and materialists believe), there may still be a
strong likelihood that we are living simulated lives. The implications of this
result are the focus of the second part of the paper. The topics discussed include:
the Doomsday argument, scepticism, the different modes of virtual life,
transcendental idealism, the Problem of Evil, and simulation ethics.
1. The Simulation Menace
Imagine participating in a simple experiment. You are watching pre-recorded scenes
from a televised soap opera unfold on a monitor in front of you; at the same time, in
different rooms, nine other people are doing likewise. Or at least, they believe they are.
In fact, only one of the screens is showing the original film featuring real actors; nine
screens are showing a computer-generated film. The simulation is very good; so good, in
fact, that the computer-generated images are visually indistinguishable from the originals.
As is clear, if you had nothing but the onscreen images to go by, then (i) you would not
be able to tell whether the people you are watching were real or computer-generated, and
(ii) the odds that you are watching the film featuring real rather than virtual people are
only one in ten.
Now consider an analogous case. As things stand, our abilities to create and
control human streams of consciousness are severely limited. Let us suppose that in the
future this changes, and it becomes possible to create human-type streams of
consciousness, of any length, with any desired characteristics, very easily. Call the
succession of streams which jointly compose the consciousness of a single person from
birth until death, a life-stream. Despite their differences, your life-stream and mine, are
of a certain general type: early 21st century human. Let us call these ‘type-21 streams’.
Now suppose that, for whatever reason, in the future very large numbers of type-21
streams will be created. To be more specific, suppose the total number of type-21
streams which exist after the year 2100 is ten times greater than the number which
existed in the 21st century itself.
This scenario places you in a similar predicament as the first, but the
consequences are rather more perturbing. Are you in a position to tell whether your
experience is real or artificially generated? No. What are the odds that your experience
is occurring when appears to be, in the early 21st century? Only one in ten. Although it
seems to you that you are a normal human being living at the start of the 21st century, the
subjects of all the many artificially produced type-21 streams have very similar
impressions and beliefs. These subjects are all mistaken, and so might you be, for it is
more likely than not that you are one of these subjects.
Following Bostrom, I will call this line of reasoning the Simulation Argument.
1
Although not everyone will find the possibility that their current lives are simulations
something to be dreaded,2
at the very least the argument threatens complacent
assumptions about the status of our lives, and for this reason I shall sometimes refer to
the simulation menace or threat. Many will no doubt be inclined to dismiss the argument
as a mildly diverting but ultimately unthreatening curiosity, and for what might seem to
be good reasons:
Could it be done? The notion that future generations, or future civilizations, will
be able to manipulate consciousness in the ways required is wildly implausible.
Would anyone bother? Even if the required technology were to become available,
our descendants would surely have better things to do than waste their time and
energy producing realistic simulations of 21st century lives, at least in the vast
numbers required for their existence to pose a significant threat to us.
In what follows I will argue that these objections carry less weight than might be
supposed. Having established that the Simulation Argument should be treated with
respect, I will move on to consider some of the implications of this.
2. Practicalities
I shall be using ‘simulation’ in a very broad way: any state or episode of consciousness is
to be regarded as simulated if it is produced by non-standard methods in a controlled
fashion (the degree of control may vary). Simulated experiences are of course real
experiences in their own right, and while a simulated episode of consciousness may be a
re-creation of an original non-simulated stretch of conscious life, it need not be. I shall
say that a life (or part of a life) is virtual rather than real if it is entirely composed of
simulated experiences.
Consciousness can be simulated in different ways, and to different degrees, and it
will prove useful to have some of these differences in view before proceeding.
So far as degree or depth of simulation is concerned, we can contrast complete
with partial simulations. The manufactured type-21 streams we encountered above areexamples of complete simulations: every part and aspect of experience is being generated
by artificial means. In partial simulations, only some parts or aspects of experience are
generated by artificial means. A simulation in which a subject is supplied with a wholly
virtual environment (which here can be taken to include all forms of bodily experience)
but retains their original psychology is one form of partial simulation. But we can also
envisage cases in which the tampering is restricted to the domain of inner experience.
Imagine having your psychology (e.g. memories, beliefs, desires, language skills,
personality traits, and so on) replaced with a replica of Napoleon’s, and then waking up
in your own bed and perceiving your environment in the usual way. In what follows,
unless otherwise stated, we will be concerned with complete rather than partial
simulations.3
As for the ways in which consciousness can be simulated, it is important to
distinguish what I will call hard (or H-simulations) from soft (or S-simulations). Hsimulations result from directly tampering with the neural hardware ordinarily
responsible for producing experience. S-simulations are streams of consciousness
generated by running programs (software) on computers (other than the brain, if the brain
is nothing but a computer).
The following scenarios outline some ways in which both H-simulations and Ssimulations might exist in menacing numbers. The scenarios may seem far-fetched (to
us, at any rate). Quite what we should make of them will be discussed in section 3.
Modal Realism and Other Worlds
According to the Modal Realist, all logically possible worlds are just as real as this one.
Suppose David Lewis is right and Modal Realism is true. Even if creating human-type
streams of consciousness in the actual world will always be a haphazard and timeconsuming business, there are many logically possible worlds where this is not the case.
There is, for example, a world which contains an infinite number of brains-in-vats, where
each brain enjoys a recognizably human range of experiences. An infinite number of
these vat-subjects, we can suppose, have menacing type-21 streams of consciousness, and
your life-stream and mine are replicated many times over. Of course, this possible
universe is not alone: there are infinitely many very much like it, differing only in the
most trivial detail. There are also many other possible universes where other methods of
simulating human-style consciousness are to be found. Suppose S-simulations are
possible; if so, there are universes where every brain-generated consciousness is
replicated by software running on computers. Suppose something like Cartesian dualism
is true, and mentality is not purely material. It makes no difference: there are
innumerable worlds, each containing innumerable disembodied immaterial souls, in
which innumerable variants of type-21 streams of consciousness are instantiated.
Needless to say, the subjects of these streams as completely unaware of their real
predicament: they believe themselves to be embodied beings, living a normal life on 21st
century Earth.
If Modal Realism were true, the simulation menace would clearly be very real
indeed. As for the odds of having a virtual rather than a real life, the calculation is not
straightforward. Since there may well be an infinite number of worlds which are close
replicas of (what we take to be) the actual world, there may well be an infinite number of
worlds containing exact (or nearly exact) replicas of (what we take to be) our 21st century
Earth. If so, then since every simulated type-21 stream could be paired off with a real
type-21 stream, it could be argued that real and simulated streams are equal in number.
In which case, the one’s odds of one’s life being real rather than simulated are at best
only fifty-fifty.
Modal Realism is not the only potential source of simulations-in-other-worlds:
some of the ‘multiverse’ theories encountered in the more speculative reaches of
contemporary physics may have similarly menacing implications. Consider, for example,
the kind of cosmos described by Smolin in which ‘each black hole is a bud that leads to a
new universe’ (1997, 94). The variants of this model which have no beginning or end in
time comprise an infinite number of sub-universes of varying character. A cosmos of this
kind could easily contain an infinite number of Earth-like planets with Earth-like
civilizations. Even if only some of the latter generate simulated type-21 streams, it could
still be possible to pair off every real type-21 stream with a simulated type-21 stream.
S-Simulations
If all logically possible worlds are real, menacing simulations in truly vast numbers will
inevitably exist. But even if only this world is real, menacing simulations may well exist
in vast numbers – especially if S-simulations are possible.
Speculations as to what computers might one day be capable of are commonplace,
but Frank Tipler takes things a good deal further than most. Tipler argues that if our
descendants develop computers as far as they can be developed, given known physical
constraints, we can all look forward to being resurrected in the far-future. Intriguingly,
he suggests that our resurrection will not depend on our descendants having detailed
knowledge of what our lives were actually like. The deduction runs thus:
- The computational conception of the mind is true. Any mental life, any stream of consciousness, can be replicated on a suitably programmed computer.
- There total number of possible human-like streams of consciousness (of finite duration) is finite.
- The processing power of the ‘universal computer’ that our descendants will develop will be effectively infinite.
- The universal computer will easily be able to simulate every possible human stream of consciousness (of finite length).4
- Hence our resurrection is all but inevitable: ‘The dead will be resurrected when the computer capacity of the universe is so large that the amount of capacity required to store all possible human simulations is an insignificant fraction of the entire capacity.’ (1994, 225)
If the future turns out as Tipler predicts, the Simulation Argument has real bite: the
probability that your life is a virtual life is extremely high. As Tipler himself notes:
‘How do we know we ourselves are not merely a simulation inside a gigantic computer?
Obviously, we can’t know.’ (1994, 207)
But Tipler’s scenario rests on a good many colossal ifs. The future he describes
may not be physically possible; even if it is, there is no reason to think it significantly
likely that the future will turn out as he describes. However, there are far more modest
forecasts which do not suffer from these flaws, but whose implications viz à vis the
simulation menace are much the same.
If computer technology continues to advance at the rate it has for the past few
decades, it will not belong before our most powerful computers equal, or exceed, the
processing power and information storage capacity of a typical human brain. According
to one of the more optimistic guesstimates, supercomputers should cross this threshold as
early as 2010, with desktop machines of similar power arriving by 2030. A more
conservative survey concludes that the breakthrough will certainly have been achieved by
supercomputers around 2025, if present trends continue – and there is every reason to
think they will.5
Once such hardware becomes available it will be possible to simulate
the computational activity of a human brain. Of course, this will require a fine-grained
knowledge of our brains’ structure and workings, but it may well be that the required
knowledge will gradually be accumulated over the next few decades; significant strides in
this direction have already been taken. On the assumption that mentality is a purely
computational affair, computerized simulations of human brains could generate conscious
mental lives that are subjectively indistinguishable from those generated by biological
brains. S-simulations of this kind could be possible within the next half century or so.
A few such simulations pose no significant threat, but the situation becomes
distinctly menacing if they start being produced by the billion or trillion. Such a situation
could develop in at least two ways. The ability to produce menacing simulations could
become very widespread, e.g., a hundred years from now everyone might own desktop
(or handheld) computers easily capable of running them. If several billion computers
were to possess this capacity, even if it were utilized only occasionally, menacing
simulations would soon exist in disturbingly large numbers.6
Alternatively, or in parallel,
the capability of running large numbers of simulations might be found in the
supercomputers of the not too far-distant future. Bostrom provides a Tipleresque
illustration of the potential dangers:
- a rough approximation of the computational power of a single planetary-mass computer is 1042 operations per second, and this assumes only already known nanotechnological designs, which are far from optimal. Such a computer could simulate the entire mental history of humankind (call this an ancestor-simulation) in less than 10-7 seconds. (2002b, 3)
Less powerful devices, of the sort which might be available in the comparatively near
future, would take somewhat longer, but the lesson remains much the same. If our
descendants were able to run ancestor-simulations using only a small fraction of the
computing resources available to them, they might very well do so, quite frequently. In
such circumstances were to obtain, the risk that you and I are inhabiting a computer
simulation would be high
H-Simulations
There is a further source of menacing simulations, one that has received less attention.
Many of us have experienced fully realistic hallucinations, whether drug-induced or in
ordinary dreams. Hallucinations produced by these means are typically uncontrolled: we
cannot determine in advance the type of virtual world we will hallucinate, or the role we
will play in the scenarios which unfold. This may very well change. Advances in brain
science may make it possible to generate controlled hallucinations – or H-simulations –
safely, easily and reliably. Almost inevitably, some of these controlled hallucinations
will constitute menacing simulations.
One route to H-simulations requires the kind of neural implant and humanmachine integration that is already familiar from science fiction. Interacting with
computers mechanically – using screens, keyboards, etc. – is a cumbersome business,
and a good deal of research is going into ways of facilitating the process. Among the
methods already being considered, at least by the more adventurous researchers, are
methods of connecting computers directly to brains. At present, such techniques are at a
primitive stage of development, but this will no doubt change.8
A hundred years from
now, children could be growing up with implants buried deep in their heads, implants that
both track and keep pace with their neural development, and allow their minds to interact
directly with computers, on a number of levels, in a variety of ways.
It is not difficult to envisage some of the uses to which this sort of interface might
be put. Your thoughts could be transmitted directly into someone else’s mind – provided
you were both hooked up to the same computer network. Forgetfulness would be largelya thing of the past: your thoughts and experiences could easily be backed-up on a
computer file, ready to be called on when required. More relevant to our purposes, fully
immersive virtual reality would also be a possibility. There will be no need for you to
wear a suit and visor to interact with machine-generated virtual worlds, your implants
will perform the necessary tasks. Your sensory experience will be directly machinecontrolled, via stimulation of the appropriate areas of the sensory cortex. The movements
of your (simulated) body through virtual environments will be under your control, but
there will be no need for you to actually move your physical body: the ways you intend to
move your body will be detected by implants in your motor cortex and elsewhere, and
this information will be used to generate corresponding movements of your virtual body.
It will be possible to have a fully realistic experience of (say) flying a plane through
narrow mountain passes while remaining motionless on a couch. You might even believe
yourself to be an experienced pilot: your implants could ensure that a suitable set of false
memories temporarily override your real memories. Alternatively, you might believe
yourself to be an ordinary 21st century person, leading a typical life in a (virtual) 21st
century environment.
If such technology were to be commonplace, it is by no means inconceivable that
H-simulations would be generated in sufficient numbers so as to become menacing.
People might take virtual reality ‘trips’ to the past quite frequently. They would certainly
be used on an occasional basis during history lessons, and more intensively by historians,
amateur and professional, with a particular interest in what it was like to live during
certain periods of the past. But such trips might also be taken – far more frequently – for
entertainment purposes. The soap operas of the future might well have an
immersive/interactive character their present-day counterparts lack, computer games
likewise. As is easily shown, the numbers soon add up.
Our descendants may ‘visit’ the past quite frequently, but since few are likely to
want to spend significant portions of their lives in H-simulations, the concept of a lifestream introduced earlier is no longer appropriate as a basic unit of simulated
consciousness. Something of briefer duration is required.9
So, for present purposes, let
us take day-long streams of uninterrupted consciousness – D-streams for short – as our
working units. (An even shorter unit could be selected, but as will become evident, the
upshot would not be greatly different.) We shall take as our class of menacing D-streams
those simulated streams that resemble the sort of experiences enjoyed by actual
inhabitants of the year 2002 – call these MD-streams.
Assuming the current population of the Earth to be six billion, there are just over
2 x 1012 D-streams for the year 2002. If a similar number of MD-streams of varying
character exist in the future, the odds of the experiences you are currently having being
simulated rather than original are around fifty per cent. Should the numbers of MDstreams created in the future be greater, your chances of living among the original
inhabitants of the year 2002 will be correspondingly smaller.
In fact, the number of MD-streams created in the future could easily be far higher.
Call the time at which H-simulations become commonplace occurrences the C-threshold.
Let us suppose that from the C-threshold on, every future human being takes one virtual reality trip to the year 2002 during their lifetime, and that these trips are varied in
character. If we now suppose that human civilization lasts for ten thousand generations
after the C-threshold, and has an average population of ten billion, there will be 1.0 x 1014
MD-streams, compared with 2 x 1012 original D-streams. With fifty simulated streams
for every real stream, you have a one in fifty chance of actually being alive in the year
2002. On more optimistic scenarios, your predicament is even more precarious. If
humankind has a long history – one million generations exist after the C-threshold, say,
with constant or improving technology – and a larger average population during this
period – a hundred billion, say – then we can expect a total of around 1.0 x 1017 MDstreams to occur, which would reduce your chances of being alive in 2002 to around one
in fifty thousand! In this case, even if only one in a thousand people ever take a virtual
reality trip back to 2002, the chances that you are really living in 2002 are still only one
in fifty.
10
3 Assessments
Of the various ways in which a simulation menace might arise, Modal Realism is the
most solid. It requires no second-guessing as to what is scientifically possible, or how
the future might turn out, and the realm of logical possibility is so vast, so unconstrained,
that menacing simulations are in inevitably in plentiful supply. But of course, Modal
Realism is itself a highly controversial metaphysical doctrine, and few philosophers
believe it to be true. That the doctrine entails a significant simulation menace will, for
many, count as yet another reason for rejecting Modal Realism.
Somewhat similar considerations apply to scenarios involving multiverses of the
menacing variety (i.e., those in which are such that it is reasonable to suppose that a
significant proportion of all conscious lives are simulated).11 Since the relevant physics
is highly speculative and controversial, it is impossible at present to know how seriously
theories of this type should be regarded. Nonetheless, the possibility that our universe
has the character these theories predict cannot be ruled out, and the same can be said of
the consequent simulation menace.
The issues raised by S-simulations are rather different. Those who have grown
familiar with the claim that the human brain is the most complex object in the known
universe may be surprised to discover that it will not be very long before we are able to
construct machines of comparable complexity and computational power. But even if this
is the case – and I suspect it is – the simulation menace posed by advances in computer
technology is less severe than Bostrom and Tipler would have us believe. For Ssimulations to constitute a threat they would have to be truly conscious. It is by no means
certain that they would beTipler takes a functionalist-cum-computationalist view of the mind as a given
(1994, 124-7). If having a mind involves nothing more than possessing the right kind of
causal organization, then in principle at least, minds can be implemented in physical (or
non-physical) systems of radically different kinds, computers included. Given the way
functionalists define mental states and properties – in entirely causal-functional terms –
computers which replicate the causal-functional organization of a human mind simply
cannot fail to be conscious. While this view certainly has its advocates, it also has many
detractors. The reasons for this are well-known and familiar. Functionalism entirely
overlooks, or ignores, the intrinsic qualitative features of experience – the very features
which provide experiences with their experiential nature! Confronted with this objection,
functionalists argue that the qualitative dimension of consciousness is irrelevant or
illusory, but in the opinion of many, myself included, these heroic arguments are entirely
unpersuasive. If functionalism is false, there is no guarantee whatsoever that computerbased simulations of human minds would be conscious.
This said, rejecting functionalism does not entirely eliminate the possibility of
computer-based consciousness. Classical Cartesian dualism offers perhaps the most
secure defence against this possibility, but another form of dualism offers no defence
whatsoever. According to the doctrine of ‘non-reductive’ or ‘dualist’ functionalism,
sympathetically explored by David Chalmers, experiences are non-physical, but they are
nomologically correlated with functional organization; on this view, computer
simulations of your brain would generate streams of consciousness indistinguishable
from your own.12 But as Chalmers himself would concede, non-reductive functionalism
is at best a possible solution to the matter-consciousness problem. Materialism is another
option, one that remains very much alive. Perhaps phenomenal properties are simply a
certain kind of material property. If so, then it may very well be that human-type
consciousness requires a human-type brain, or at least a biological system of a similar
kind.13 Of course, we cannot be certain of this. We do not know which parts or aspects
of the physical processes in our brains are responsible for producing consciousness;
consequently, we cannot rule out the possibility that the relevant physical processes could
be replicated in very different physical systems – perhaps even silicon chips. But this is
no more than a possibility, and in all likelihood, one that is quite remote if materialism is
true.
So far as the computational challenge is concerned, the situation is clearly far
from clear-cut. Whereas functionalists – of both classical and dualist persuasions – have
good reasons for being very wary of future developments in computer technology, those
who subscribe to different options in the philosophy of mind have far less reason to feelgreatly concerned: brain-simulations that are not truly conscious pose no menace
whatsoever.
By contrast, the threat posed by H-simulations looks to be very real indeed,
irrespective of the position one adopts in the philosophy of mind. The simulations
generated by the processes falling under this heading are indisputably streams of
consciousness, rather than mere computation. Since the envisaged streams are generated
by brains, rather than computers, even the most biologically oriented of materialists has
reason to take their possibility seriously.
A materialist might argue that brain-computer connections, of the kind required
for the scenario sketched above, would be so invasive and pervasive that their presence
inside a brain would be incompatible with the production of conscious experience. Given
our ignorance of the physical processes underlying consciousness, this possibility cannot
be ruled out, but there is also little reason to suppose it very likely. After all, the
envisaged interfaces would not replace neurons as experience-producers, they would
merely provide ways of artificially controlling the triggering of neurons, or neural circuits
– and we already know this to be possible on a small scale. Needless to say, the required
nano-scale technology is far beyond anything we are capable of producing at present, and
even if we could, our understanding of the brain’s functioning is not sufficiently
advanced for it to be deployed effectively. But anyone inclined to think this will continue
to be the case should bear in mind Arthur C. Clarke’s well-known dictum that any
sufficiently advanced technology is indistinguishable from magic. A hundred years ago,
now-routine procedures such as organ transplants and genetic engineering would have
seemed miraculous. Is it not probable that a hundred years from now our descendants
will be capable of similarly impressive feats?
It is not only materialists who should be open to the possibility of H-simulations,
dualists should be too.
Even if our experiences unfold within immaterial substances, it is evident that our
minds are profoundly dependent upon our brains. No contemporary dualist would be
inclined to deny that the course of our sensory experience is dependent upon the neural
activity within our brains, and this fact alone opens up the possibility of controlled
hallucinations of a limited kind. But dualists should also recognize that appropriate
neural manipulation could impact upon our conscious beliefs, intentions and desires.
Intoxicants do not merely make it harder to control our bodily movements, they make it
harder to think clearly, and there are numerous forms of brain damage that have more farreaching (and often permanent) effects on our personalities and cognitive functioning,
memory included. If brain damage can result in the permanent loss of certain memories,
is it not likely that the memories to which we have conscious access depend on
information stored in our brains? In which case, appropriate neural manipulation could
lead a 23rd century person to have access to apparent-memories of the sort a 21st century
person would have had.
But there is a further point to note, one that is relevant to materialists as well as
dualists. Brain-computer interfaces of the kind I have been considering offer the
possibility of very tightly controlled hallucinations, but there are undoubtedly other ways
of inducing similarly life-like H-simulations, even if they offer rather less potential for
fine-grained control. Ordinary, unaugmented, human minds are able to fashion richlydetailed and real-seeming virtual realities all on their own, almost effortlessly. Ordinarydreams provide evidence both of this, and our ability to spin complex virtual worlds from
limited and/or fragmentary evidence. I expect most of us have found ourselves having
vivid dreams set in (say) the 17th century shortly after watching a film set in the same
period. Although the dreamed-environment in such cases is inspired by what was seen
onscreen, it often has a depth and complexity all of its own. Future methods of
experience-induction could easily exploit these ordinary abilities. All that would be
required is a safe and reliable drug which enabled people to enter a dream-like state at
will, and also direct the general direction of the subsequent (fully life-like) hallucination
– the framework for the latter could be supplied by a little prior reading, or the watching
of video footage (e.g., of a 21st century televised soap opera). This method of controlling
hallucinations could be put to the same uses – in, say, education and entertainment – as
the computer-driven variant we considered earlier, and so is likely to be widely
employed. So far as I can see, this method of inducing (partially) controlled
hallucinations is not ruled out by any philosophical conception of the mind. It is also
quite likely to prove attainable, perhaps quite soon.
4. Consequences
Modal realism is metaphysically extravagant, developments in physics may lead to the
Many-Worlds hypothesis’ being discarded, and the computational conception of mind is
highly suspect. But while those who reject these doctrines can take comfort in the idea
that in so doing they are greatly diminishing the simulation menace, they would be wrong
to conclude that the latter is non-existent. As we have seen, there are other ways of
producing menacing simulations, ways that are harder to rule out. As one might expect,
acknowledging the force of the Simulation Argument has a number of consequences; I
will briefly outline a few of the more significant and intriguing.
WHERE DO WE STAND?
The simulation menace may be real, but it would be premature to suppose that we are all
living virtual rather than real lives; after all, we do not know how the future will turn out.
What the Simulation Argument does reveal, on the face of it at least, is a tension between
the following propositions:
- Humankind will have a long and successful future.
- Technology will make it possible to manufacture realistic experiential simulations of any known type of human life, and these will be created frequently, in many and varied forms.
- You and I exist in the early 21st century.
4.Modal Realism is true and/or a (menacing form of) multiverse theory is true.
Again, the more confident you are that (4) is true, the less confident you should be that
(3) is true
Comments
Post a Comment