How the Hard Problem Matters

Summary. While sometimes seen as an abstruse philosophical concern, the hard problem of consciousness raises questions of enormous practical importance. A deeper understanding of how phenomenal experience works on a fundamental level would help in determining whether entities like insects, advanced computers, or even rocks can feel pain and pleasure. (See Note below.)

Note: What appears below the next horizontal divider is an essay consisting of a series of questions that I've since realized are confused. Unless we take dualism seriously -- and I think Occam's razor strongly militates against doing so -- then it doesn't make sense to ask questions like, Does a computer simulation of a mind really instantiate consciousness? That question reifies consciousness as a thing that may or may not be produced by a simulation, which presupposes a separate consciousness stuff beyond the particles making up the computer on which the simulation is running. Rather, the particles constituting the computer just move -- and that's it. The question of whether a given physical operation is conscious is not a factual dispute but a definitional one: Do we want to define consciousness as including those sorts of physical operations?

What utilitarians need to decide, then, is Do I want to care about such and such particle movements similarly to the way in which I care about the particle movements that happen when, say, my own brain runs a similar algorithm on its own wetware? All of the questions asked in my original essay can make sense once again if we replace the word consciousness with physical operations that I decide to care about because the underlying algorithm is similar enough to the one in my own head that produces self-awareness and feel pain by physical operations sufficiently similar to those that I call 'pain' in myself that I decide they're something I want to prevent.

Here are the answers to some of the questions posed in the original piece. I claim no originality to any of these ideas -- they represent the standard position of the reductionists like Daniel Dennett who solve the hard problem by dissolving it.

• The hard problem: Why does it feel like anything to have the type of self-awareness that humans do? Answer: When the particles that compose self-aware brains like ours move around, they do so in such a way that the parts of the brain that perform higher-level executive functions (what we call the conscious parts) receive signals of a type that cause them to move in such a way that they, among other things, transmit signals to the mouth to utter the words, That feels like something. Qualia feel like something because the organisms that experience them execute cognitive algorithms that make them act in ways that we call believing that it feels like something. I'll put it another way: If you hold an implicitly dualist view of consciousness, what do you think happens mechanically in the brain when an organism feels qualia? Well, those mechanical operations are what qualia are. Qualia are labels we use to divide up thing-space into categories; they don't correspond to something physically real. Why we have them and how they work in the brain are interesting questions, but they're not mysterious -- they're akin to the question, Why does Lisp have gensyms and how are they implemented (both at the level of source code and at the level of hardware)?

• Why aren't we philosophical zombies? Answer: We are zombies!

• Do insects feel pain? I'll rephrase the question: Are the physical operations that go on inside insect brains sufficiently similar to the operations in our brains which we call 'suffering' that we decide we want to care about them? Fundamentally, it's up to us. Still, the question isn't a trivial one. What we need to do is decide upon some set of criteria for whether a given physical process is of the type that we want to regard as being a mind that we care about. For example, we may decide to regard as bad those mechanical operations which correspond to the execution of an algorithm for self-awareness of painful input signals (where "self-awareness" may have a very sophisticated definition, if we so choose). Whether insects do this type of self-modeling of their reception of emotional signals (or whether they receive the emotional signals without a higher-level understanding of what's going on) is then an open factual question -- and a perfectly understandable one. One way to begin crafting that set of criteria for what deserves to be called conscious suffering would be to examine the neural correlates of suffering in humans and other complex animals, and try to generalize what it is about those particular brain algorithms or processes that we can in general say is bad.

• In the previous response, you used the phrase, mechanical operations which correspond to an algorithm. But we can't uniquely decode physical processes as corresponding to, say, a particular piece of source code. There are lots of joke interpretations we could use. By a cleverly rigged correspondence relationship, we could 'call' pretty much any physical operation the execution of a suffering algorithm. This is true, and it is worth contemplating. However, it's nothing mysterious. Physics is physics -- particles move around. Stuff happens. It's up to us how we want to respond to particular sorts of particle movements. One proposal I've heard is to care about all physical operations with a different measure -- that is, in proportion to how naturally they can be interpreted as the algorithm we care about. (For more on a measure approach to consciousness, see the writings of Paul Almond.) This measure is not physical like the measure proposed by the many-worlds interpretation of quantum mechanics -- just a degree of caring, something totally non-mysterious. Since consciousness simply refers to a cluster of physical operations within thing-space, we can decide to use a fuzzy logic-style approach in saying that some things are closer to what we mean by conscious than others.

• Is a rock conscious? Answer: Do you want to care about the atomic movements in a rock as though they're conscious? I personally attach little concern to the particle movements in a rock. In view of the previous question, I may care to a non-zero degree that certain strained interpretations of rock particle movements correspond to certain algorithms that, running in animal minds, constitute self-awareness of suffering, but this degree of concern is tiny -- probably small enough to ignore entirely from my calculations.

• If we write a simple Python program with an Organism class that has a field for Pleasure and set it equal to the string infinity, have we brought an infinite amount of pleasure into the multiverse? Answer: It depends on whether you want to regard such a program as having such infinite value. I don't. In fact, I may very well regard it as having no value.

• Is a simulation of my brain on a computer conscious just like I am? Answer: Do you want to care about a computer simulation of your brain in the way you care about your brain? It's up to you. I'm not entirely sure for myself; I may be something of a carbon chauvinist in caring more about those algorithms when they're executed by the atoms of my neurons rather than the atoms of a silicon duplicate. My degree of caring may not be substrate-neutral, though the fading qualia thought experiment gives me strong contrary intuition that it ought to be.

• Is degree of pain proportional to the amount of neural matter involved in generating it? Answer: Do we want to care about algorithms run on a greater amount of hardware more than those same algorithms run on less hardware? I think I probably do, especially in light of Nick Bostrom's thought experiments, but I'm not entirely sure -- my intuitions are still fuzzy.

  Proportioning concern based on brain size may be a rough heuristic, but I think we should be wary of extending it too far. For instance, suppose certain insects do run algorithms that self-model their own reaction to pain signals in a similar way to how this happens in humans. (See, for instance, pp. 77-81 of Jeffrey A. Lockwood, The Moral Standing of Insects and the Ethics of Extinction for a summary of evidence favoring insect consciousness.) If bees have only 950,000 neurons while humans have 100 billion, should we count human suffering exactly 105,263 times as much as comparable bee suffering? I would argue not, for a few reasons.

  • The relevant figures should not be the number or mass of neurons in the whole brain but only in those parts of the brain that run the given pieces of code. I would guess that human brains contain a lot more executable code corresponding to assorted brain functions that bees lack than vice versa, so the numbers of relevant neurons are probably not nearly as different as naive division suggests.
  • The kludgey human brain, presumably containing significant amounts of legacy code, is probably a lot bulkier than the more highly optimized bee cognitive architecture. This is no doubt partly because evolution constrained bee brains to run on small amounts of hardware and with low power requirements, in contrast to what massive human brains can do, powered by high-calorie consumption and an endothermic metabolism. Think of the design differences between an operating system for, say, a hearing aid versus a supercomputer. If we care more about the number of instances of an algorithm that are run more than, e.g., the number of CPU instructions executed, the difference between bees and humans shrinks further.

  These points raise some important general questions: How much extra weight (if any) should we give to brains that contain lots of extra features that aren't used? For instance, if we cared about the number of hearing-aid audio-processing algorithms run, would it matter if the same high-level algorithm were executed on a device using the ADRO operating system versus a high-performance computer running Microsoft Vista? What about an algorithm that uses quicksort vs. one using bubblesort? Obviously these are just computer analogies to what are probably very different wetware operations in biological brains, but the underlying concepts remain. (I should add that current computers don't presently run self-modeling algorithms anywhere near similar enough to those of conscious animals for me to extend them moral concern.)

A friend posed some questions regarding my above position. Below I present some snippets from that exchange.

• Q: It seems completely arbitrary to me to show concern for certain movements of particles and not for others.

  A: Arbitrary, yes. But that's the way I'm hard-wired: I'm a machine that responds to thoughts about suffering (i.e., certain particle movements) with a desire to act to reduce it. All machines respond to certain types of particle movements -- that's how machines work. Indeed, how could things be otherwise? What else could it possibly look like for a being to care about something and respond to it other than to care about real things in the world and take physical actions in response?

  Suppose there were some sort of pain particle corresponding to the qualia of suffering. Why care about that? What makes that any less arbitrary than a particular class of particle movements corresponding to particular cognitive algorithms within certain sorts of self-aware brains?

  Maybe you would say it is indeed arbitrary to care about certain kinds of physical particles and not others; however, pain is not physical but somehow other-worldly or of a different metaphysical type. I'm not sure what it means to say that something which interacts with the physical world is other-worldly, but okay, sure. There's still the question, Why care about other-worldly things? Why care about things with different metaphysical types? Suppose the only thing that had a different metaphysical type than physical things was the feeling you get when you put your foot into silly putty. Would that then be all that you care about?

• Q: It remains completely unexplained why you do in fact care about such patterns and not others. Also, it doesn't explain why, once you realize that there is nothing special about them, you still care about them.

  A: Say you enjoy eating ice cream. Well, ice cream is just a particular collection of a subset of the atoms from the periodic table arranged in a particular way. It's completely unexplained (in the sense of justification, not evolutionary origins) why you enjoy that group of atoms and not arrangements of, say, iron oxide -- how do you justify that fact? And does your failure to do so make it taste less good?

  This just restates the point that all ethics is ultimately arbitrary -- it's we who have to decide what we want to care about. I don't see caring about a certain class of algorithms instatiated by physical brains as less arbitrary than caring about anything else. If it helps, we can think about this class as being special -- whatever that means. We can have a sense that there's a unique, mysterious thing that it's like to be in pain, and indeed, that language describes what we mean by our caring. But that is, basically, poetry -- it shouldn't prevent us from understanding physically what's really going on.

  Talking in a distanced way about physical mechanisms is important, to ensure that we have a correct understanding of the physical world. However, it shouldn't lead us to see emotions coldly or lose motivation to prevent suffering. There are many levels of abstraction for describing a process, each useful for its own purposes. The more poetic description of self-awareness of certain neural processes that we call feelings is entirely appropriate as well. We can also remind ourselves of the seriousness of suffering by raw experience: Does pain hurt any less once you've learned that it's produced by chemical and electrical signals transmitted via nerve cells? Knowing how something happens doesn't make me stop caring about it.

• Q: Finally, it is unclear to me how, once you acknowledge that your concern for pain is ultimately arbitrary, can you draw a fundamental difference between concern for pain and the innumerable other concerns that we have pre-reflectively. If all these concerns are on a par, it seems very unlikely that only one of them would survive a process of critical examination while all others would eventually dissolve or disappear. Again, it seems that we must recognize pain as a special type of concern to explain the fact that it is the only one that remains.

  A: Unexplained, yes. But that's the way I'm built. Not only do I care about lots of particle movements (preventing suffering, eating ice cream, sleeping when tired, etc.) but not others (e.g., creating paperclips, pushing sand around in circles, sorting pebbles into heaps), I also have a special regard for one of those ways of moving particles, namely preventing suffering: I (claim to) place it far above the others in importance. (As a practical matter, it's not clear how much that's the case: I am also a somewhat selfish human being who often fails to spend my resources in a completely optimal manner.) The fact that this machine thinks about certain particle movements as belonging to a unique class is just another of the machine's interesting attributes.

  A good way to summarize is as follows. At an intuitive level, I still do think in precisely the terms you describe: A what it's like to feel pain that's specially bad and whose alleviation ought to take priority over other things people value. But how the cognitive algorithm feels from the inside shouldn't get in the way of our really understanding it. One reason it's important to look at the algorithm from the outside as well is that doing so helps us avoid, for instance, the mistake that our AIs will automatically be friendly or empathetic.

The Importance of Qualia

Utilitarians, and anyone who cares about reducing suffering, are fundamentally concerned with qualia. Our ultimate goal is to prevent experiences of distressful qualia and create more feelings of pleasant qualia. A natural question that arises is, What sorts of entities experience these qualia? In particular, below are a few instances of questions of this sort whose answers have significant ramifications for the types of policies utilitarians should support and the causes on which they should expend resources.

1. Which animals can suffer? How far down the evolutionary tree does conscious awareness of pain extend? In particular, can insects suffer? Fortunately, these questions do not require a complete understanding of consciousness before we can arrive at good answers. I'm already, say, 97% sure that you can feel pain, because you appear to be an organism with nearly identical physiology to my own, behaving in very similar ways. Similarly, I'm 95% certain that cats can suffer for similar reasons, including shared phylogenetic heritage, comparability of neural structures, and similarity of behavior under distress. However, when it comes to fish, I would assign only, say, 70% probability to the capacity for conscious suffering, and insects are considerably lower than that.

2. Can non-animals suffer? One view of consciousness is panpsychism, which holds that all matter in the universe has mind stuff that gives rise to phenomenal experience. If this is true, at what level do conscious minds begin? I as a whole organism feel conscious, but what about each of my neurons? Should I worry about rocks being conscious? Even if so, the utilitarian implications wouldn't be obvious -- Do rocks prefer to lie in this pile or that one? -- but it's possible that further reflection on the matter would suggest nontrivial recommendations for action.

3. Can artificial computers be made conscious? This is relevant for a number of reasons.

   • If humans do create conscious computers, we will need need to enforce computer welfare guidelines in the same way that we impose animal-welfare standards. (Of course, there may be utilitarian reasons for subjecting conscious computers to various operations, just as there are often utilitarian reasons for animal experimentation.)

   • Digital computers are highly efficient at certain types of computations (e.g., adding numbers), and it's conceivable that they could be made efficient at computing conscious experiences. Applying the idea of computronium to the production of utility, we could try to fill the universe with utilitronium, or at the very least, with astronomical numbers of simulated human lives. (However, concerns about creating large numbers of damnable souls need to be considered.)

   • The answer to the question affects the probability we assign to Nick Bostrom's simulation argument, since if computers can't produce conscious experiences, then in order to be simulations, we would presumably have to be running on biological substrates, which are presumably harder to create in large numbers. Whether we are simulations has an impact on, among other things, our approach to religion and immortality. For instance, we might appeal to our simulators for miracles. Moreover, the universe in which our simulators live might have different laws of physics, perhaps including infinite computational power.

The Hard Problem

For the first of these questions, Which animals can suffer?, one very promising approach is to identify the neural correlates of consciousness (NCCs) in humans. Even if we don't understand the causes of consciousness, we can get a pretty good sense of what sorts of neural activity are associated with consciousness and then check whether activity of this type occurs in other animals.

Of course, as we move further backwards in the evolutionary tree, using this criterion alone becomes increasingly dubious: Do animals need to have the specific neural structures that humans do in order to feel qualia? Might they not have phenomenal experience through other biological mechanisms? And when it comes to the second and third questions, the NCC approach is almost entirely inapplicable. Rather than merely identifying the neural correlates of consciousness, utilitarians really want to know the correlates of consciousness in general: What necessary and sufficient criteria would allow us to tell that entity X is suffering while entity Y isn't?

A second approach to assessing the presence of consciousness, familiar from the animal-welfare literature, is to examine behavior: Does this organism, when given a noxious stimulus, exhibit prolonged aversive responses? Does the stimulus affect learning and motivational tradeoffs? Is it remembered? How similar are the behavioral reactions to humans'? In general, it seems safe to assume that consciousness is somewhat widespread throughout the animal kingdom, because it exists in Homo sapiens and appears to be a useful adaptation. Or is it? This raises the question: What exactly is consciousness for? Sure, perhaps a system for reflective awareness allows an organism to organize its thoughts, imagine counterfactual scenarios, and execute sophisticated novel behaviors. But why does the process feel like anything? Why doesn't it take place in the dark, to quote David Chalmers.

This is precisely the hard problem of consciousness that Chalmers has promulgated: It is widely agreed that experience arises from a physical basis, but we have no good explanation of why and how it so arises. Why should physical processing give rise to a rich inner life at all? Chalmers himself proposes that consciousness may be fundamental to reality, in the same way that, say, general relativity takes space-time as a foundational building block that is simply assumed to exist. On this property-dualist view, physical entities interact with mental entities in a structurally coherent way, i.e., any information that is consciously experienced will also be cognitively represented. Others have criticized this position, and needless to say, a plethora of alternate views have been proposed regarding the philosophy of mind.

Some of these disputes are more metaphysical than I care to think about. For instance, Does consciousness exist as a separate entity, or is it just one concept-space partition within underlying reality that we can choose to describe? I take more of an engineer's view, interested in the question: What do I need to do in order to identify painful experiences (and then hopefully prevent some of them)? Of course, just as aerospace engineers designing GPS systems can't neglect relativity, utilitarian engineers need some understanding of the underlying philosophical and scientific theory of consciousness in order to determine, say, whether they need to worry about causing pain through computer simulations. In this sense, utilitarianism tells us why we care about the hard problem: We need to know enough about the conditions that give rise to qualia to determine how we should act toward entities like animals and computers that can't self-report this information. (A computer that displays the statement I'm conscious on its screen isn't convincing....)

Conscious Simulations?

The important part of the hard problem, then, is not so much why we feel qualia. As John F. Kihlstrom noted (Scientific Approaches to Consciousness, Spring 2009, Lecture 6), that's sort of like asking, Daddy, why is there air? (which we might call the hard problem of air). What we really need to know is how phenomenal experiences are produced, so that we can create more good ones and fewer bad ones. Still, the explanatory gap that concerns Chalmers and others isn't completely trivial, either. It's not enough just to solve the easy problems of consciousness, like how we react to our environment, focus our attention, or make decisions, because, for instance, robots can do those same sorts of things, and it's not clear at what features the robot has to have before it begins consciously to experience the world. Barring panpsychism, a thermostat is probably not conscious, while a neuron-for-neuron replication of my brain may very well be. Where's the dividing line?

And even if we reject carbon chauvinism toward consciousness, is it the case that the physical substrate underlying computation doesn't matter at all? Could an arbitrary universal turing machine give rise to qualia? Does that include a man manipulating Chinese symbols, or a Lego computer? What about a mechanical device performing a single Turing-machine operation every million years? What about physical processes that just happen to be interpretable as implementing a consciousness-producing algorithm -- what Gary Drescher calls joke interpretations of consciousness (Good and Real, pp. 54-55)? Certainly not a giant lookup table?

If qualia are merely computational, what sorts of internal representations do they require? Could one, as Eliezer Yudkowsky asks, redesign[...] the brain to represent the intensity of pleasure using IEEE 754 double-precision floating-point numbers, [so that] a mere 64 bits would suffice to feel pleasures up to 10^308 hedons [...]? What happens when we introduce +/- INF symbols?^[1] My guess is that this suggestion wouldn't work: That, even if qualia are generated by computations on an arbitrary physical instantiation of an appropriate Turing machine, you still need to execute one iteration of the produce-the-feeling-of-happiness algorithm in order to get a hedon out of it. It's not enough, I assume, to talk about running the algorithm lots of times, just as it's not the case that the mere existence of a given Turing machine in the Platonic mathematical realm is sufficient for that computation to actually take place.

Even if simulations can produce qualia when executed, what counts as an execution? How do we avoid John Searle's suggestion (The Rediscovery of Mind, 1992) that

the wall behind my back is right now implementing the Worstar program, because there is some pattern of molecule movements that is isomorphic with the formal structure of Wordstar. But if the wall is implementing Wordstar, if it is a big enough wall it is implementing any program, including any program implemented in the brain.

Is there a difference between running a brain directly in silico versus executing a simulation of a universe in which the laws of physics happen to include the physical events that constitute a silicon brain (or a biological brain, for that matter)? Does calculating mathematically the atomic-level movements of particles making up a pain nerve generate the same experience as actually making that nerve fire in the real world? I'm skeptical, but remain open to being convinced. This question is relevant because it determines whether we ought to be concerned about AIs that might be running vast numbers of physical simulations of various universes for scientific or game-theoretic purposes.