Reinstalling Eden

If someday we can create consciousness inside computers, what moral obligations will we have to the conscious beings we create?

R. Scott Bakker and I have written a short story about this, which came out today in Nature.

You might think that it would be a huge moral triumph to create a society of millions of actually conscious, happy beings inside one's computer, who think they are living, peacefully and comfortably, in the base level of reality -- Eden, but better! Divinity done right!

On the other hand, there might be something creepy and problematic about playing God in that way. Arguably, such creatures should be given self-knowledge, autonomy, and control over their own world -- but then we might end up, again, with evil, or even with an entity both intellectually superior to us and hostile.

[For Scott's and my first go-round on these issues, see here.]

Introspecting My Visual Experience "as of" Seeing a Hat?

In "The Unreliability of Naive Introspection" (here and here), I argue, contra a philosophical tradition going back at least to Descartes, that we have much better knowledge of middle-sized objects in the world around us than we do of our stream of sensory experience while perceiving those objects.

As I write near the end of that paper:

The tomato is stable. My visual experience as I look at the tomato shifts with each saccade, each blink, each observation of a blemish, each alteration of attention, with the adaptation of my eyes to lighting and color. My thoughts, my images, my itches, my pains – all bound away as I think about them, or remain only as interrupted, theatrical versions of themselves. Nor can I hold them still even as artificial specimens – as I reflect on one aspect of the experience, it alters and grows, or it crumbles. The unattended aspects undergo their own changes too. If outward things were so evasive, they’d also mystify and mislead.

Last Saturday, I defended this view for three hours before commentator Carlotta Pavese and a number of other New York philosophers (including Ned Block, Paul Boghossian, David Chalmers, Paul Horwich, Chris Peacocke, Jim Pryor).

One question -- raised first, I think, by Paul B. then later by Jim -- was this: Don't I know that I'm having a visual experience as of seeing a hat at least as well as I know that there is in fact a real hat in front of me? I could be wrong about the hat without being wrong about the visual experience as of seeing a hat, but to be wrong about having a visual experience as of seeing a hat, well, maybe it's not impossible but at least it's a weird, unusual case.

I was a bit rustier in answering this question than I would have been in 2009 -- partly, I suspect, because I never articulated in writing my standard response to that concern. So let me do so now.

First, we need to know what kind of mental state this is about which I supposedly have excellent knowledge. Here's one possibility: To have "a visual experience as of seeing a hat" is to have a visual experience of the type that is normally caused by seeing hats. In other words, when I judge that I'm having this experience, I'm making a causal generalization about the normal origins of experiences of the present type. But it seems doubtful that I know better what types of visual experiences normally arise in the course of seeing hats than I know that there is a hat in front of me. In any case, such causal generalizations are not sort of thing defenders of introspection usually have in mind.

Here's another interpretative possibility: In judging that I am having a visual experience as of seeing a hat, I am reporting an inclination to reach a certain judgment. I am reporting an inclination to judge that there is a hat in front of me, and I am reporting that that inclination is somehow caused by or grounded in my current visual experience. On this reading of the claim, what I am accurate about is that I have a certain attitude -- an inclination to judge. But attitudes are not conscious experiences. Inclinations to judge are one thing; visual experiences another. I might be very accurate in my judgment that I am inclined to reach a certain judgment about the world (and on such-and-such grounds), but that's not knowledge of my stream of sensory experience.

(In a couple of other essays, I discuss self-knowledge of attitudes. I argue that our self-knowledge of our judgments is pretty good when the matter is of little importance to our self-conception and when the tendency to verbally espouse the content of the judgment is central to the dispositional syndrome constitutive of reaching that judgment. Excellent knowledge of such partially self-fulfilling attitudes is quite a different matter from excellent knowledge of the stream of experience.)

So how about this interpretative possibility? To say I know that I am having a visual experience as of seeing a hat is to say that I am having a visual experience with such-and-such specific phenomenal features, e.g., this-shade-here, this-shape-here, this-piece-of-representational-content-there, and maybe this-holistic-character. If we're careful to read such judgments purely as judgments about features of my current stream of visual experience, I see no reason to think we would be highly trustworthy in them. Such structural features of the stream of experience are exactly the kinds of things about which I've argued we are apt to err: what it's like to see a tilted coin at an oblique angle, how fast color and shape experience get hazy toward the periphery, how stable or shifty the phenomenology of shape and color is, how richly penetrated visual experience is with cognitive content. These are topics of confusion and dispute in philosophy and consciousness studies, not matters we introspect with near infallibility.

Part of the issue here, I think, is that certain mental states have both a phenomenal face and a functional face. When I judge that I see something or that I'm hungry or that I want something, I am typically reaching a judgment that is in part about my stream of conscious experience and in part about my physiology, dispositions, and causal position in the world. If we think carefully about even medium-sized features of the phenomenological face of such hybrid mental states -- about what, exactly, it's like to experience hunger (how far does it spread in subjective bodily space, how much is it like a twisting or pressure or pain or...?) or about what, exactly, it's like to see a hat (how stable is that experience, how rich with detail, how do I experience the hat's non-canonical perspective...?), we quickly reach the limits of introspective reliability. My judgments about even medium-sized features of my visual experience are dubious. But I can easily answer a whole range of questions about comparably medium-sized features of the hat itself (its braiding, where the stitches are, its size and stability and solidity).

Update, November 25 [revised 5:24 pm]:

Paul Boghossian writes:

I haven't had a chance to think carefully about what you say, but I wanted to clarify the point I was making, which wasn't quite what you say on the blog, that it would be a weird, unusual case in which one misdescribes one's own perceptual states.

I was imagining that one was given the task of carefully describing the surface of a table and giving a very attentive description full of detail of the whorls here and the color there. One then discovers that all along one has just been a brain in a vat being fed experiences. At that point, it would be very natural to conclude that one had been merely describing the visual images that one had enjoyed as opposed to any table. Since one can so easily retreat from saying that one had been describing a table to saying that one had been describing one's mental image of a table, it's hard to see how one could be much better at the former than at the latter.

Roger White then made the same point without using the brain in a vat scenario.

I do feel some sympathy for the thought that you get something right in such a case -- but what exactly you get right, and how dependably... well, that's the tricky issue!

Skepticism, Godzilla, and the Artificial Computerized Many-Branching You

Nick Bostrom has argued that we might be sims. A technologically advanced society might use hugely powerful computers, he says, to run "ancestor simulations" containing actually conscious people who think they are living, say, on Earth in the early 21st century but who in fact live entirely inside an advanced computational system. David Chalmers has considered a similar possibility in his well-known commentary on the movie The Matrix.

Neither Bostrom nor Chalmers is inclined to draw skeptical conclusions from this possibility. If we are living in a giant sim, they suggest, that sim is simply our reality: All the people we know still exist (they're sims just like us) and the objects we interact with still exist (fundamentally constructed from computational resources, but still predictable, manipulable, interactive with other such objects, and experienced by us in all their sensory glory). However, it seems quite possible to me that if we are living in a sim, it might well be a small sim -- one run by a child, say, for entertainment. We might live for three hours' time on a game clock, existing mainly as citizens who will give entertaining reactions when, to their surprise, Godzilla tromps through. Or it might be just me and my computer and my room, in an hour-long sim run by a scientist interested in human cognition about philosophical problems.

Bostrom has responded that to really evaluate the case we need a better sense of what are more likely vs. less likely simulation scenarios. One large-sim-friendly thought is this: Maybe the most efficient way to create simulated people is to evolve up a large scale society over a long period of (sim-clock) time. Another is this: Maybe we should expect a technologically advanced society capable of running sims to have enforceable ethical standards against running small sims that contain actually conscious people.

However, I don't see compelling reason to accept such (relatively) comfortable thoughts. Consider the possibility I will call the Many-Branching Sim.

Suppose it turns out the best way to create actually conscious simulated people is to run a whole simulated universe forward billions of years (sim-years on the simulation clock) from a Big Bang, or millions of years on an Earth plus stars, or thousands of years from the formation of human agriculture -- a large-sim scenario. And suppose that some group of researchers actually does this. Consider, now, a second group of researchers who also want to host a society of simulated people. It seems they have a choice: Either they could run a new sim from the ground up, starting at the beginning and clocking forward, or they could take a snapshot of one stage of the first group's sim and make a copy. Which would be more efficient? It's not clear: It depends on how easy it is to take and store a snapshot and implement it on another device. But on the face of it, I don't see why we ought to suppose that copying would take more time or more computational resources than evolving a sim up from ground.

Consider the 21st century game Sim City. If you want a bustling metropolis, you can either grow one from scratch or you can use one of the many copies created by the programmers or users. Or you could grow one from scratch and then save stages of it on your computer, shutting the thing down when things don't go the way you like and starting again from a save point; or you could make copied variants of the same city that grow in different directions.

The Many-Branching Sim scenario is the possibility that there is a root sim that is large and stable, starting from some point in the deep past, and then this root sim was copied into one or more branch sims that start from a save point. If there are many branch sims, it might be that I am in one of them, rather than in a root sim or a non-branching sim. Maybe one company made the root sim for Earth, took a snapshot in November 2013 on the sim clock, then sold thousands or millions of copies to researchers and computer gamers who now run short-term branch sims for whatever purposes they might have. In such a scenario, the future of the branch sim in which I am living might be rather short -- a few minutes or hours or years. The past might be conceptualized either as short or as long, depending on whether the past in the root sim counts as "this world's" past.

Issues of personal identity arise. If the snapshot of the root sim was taken at root sim clock time November 1, 2013, then the root sim contains an "Eric Schwitzgebel" who was 45 years old at the time. The branch sims would also contain many other "Eric Schwitzgebels" developing forward from that point, of which I would be one. How should I think of my relationship to those other Erics? Should I take comfort in the fact that some of them will continue on to full and interesting lives (perhaps of very different sorts) even if most of them, including probably this particular instantiation of me, now in a hotel in New York City, will soon be stopped and deleted? Or to the extent I am interested in my own future rather than merely the future of people similar to me, should I be concerned primarily about what is happening in this particular branch sim? As Godzilla steps down on me, shall I try to take comfort in the possibility that the kid running the show will delete this copy of the sim after he has enjoyed viewing the rampage, then restart from a save point with New York intact? Or would deleting this branch be the destruction of my whole world?

Expert Disagreement as a Reason for Doubt about the Metaphysics of Mind (Or: David Chalmers Exists, Therefore You Don't Know)

Probably you have some opinions about the relative merit of different metaphysical positions about the mind, such as materialism vs. dualism vs. idealism vs. alternatives that reject all three options or seek to compromise among them. Of course, no matter what your position is, there are philosophers who will disagree with you -- philosophers whom you might normally regard as your intellectual peers or even your intellectual superiors in such matters – people, that is, who would seem to be at least as well-informed and intellectually capable as you are. What should you make of that fact?

Normally, when experts disagree about some proposition, doubt about that proposition is the most reasonable response. Not always, though! Plausibly, one might disregard a group of experts if those experts are: (1.) a tiny minority; (2.) plainly much more biased than the remaining experts; (3.) much less well-informed or intelligent than the remaining experts; or (4.) committed to a view that is so obviously undeserving of credence that we can justifiably disregard anyone who espouses it. None of these four conditions seems to apply to dissent within the metaphysics of mind. (Maybe we could exclude a few minority positions for such reasons, but that will hardly resolve the issue.)

Thomas Kelly (2005) has argued that you may disregard peer dissent when you have “thoroughly scrutinized the available evidence and arguments” on which your disagreeing peer’s judgment is based. But we cannot disregard peer disagreement in philosophy of mind on the grounds that this condition is met. The condition is not met! No philosopher has thoroughly scrutinized the evidence and arguments on which all of her disagreeing peers’ views are based. The field is too large. Some philosophers are more expert on the literature on a priori metaphysics, others on arguments in the history of philosophy, others on empirical issues; and these broad literatures further divide into subliteratures and sub-subliteratures with which philosophers are differently acquainted. You might be quite well informed overall. You’ve read Jackson’s (1986) Mary argument, for example, and some of the responses to it. You have an opinion. Maybe you have a favorite objection. But unless you are a serious Mary-ologist, you won’t have read all of the objections to that argument, nor all the arguments offered against taking your favorite objection seriously. You will have epistemic peers and probably epistemic superiors whose views are based on arguments which you have not even briefly examined, much less thoroughly scrutinized.

Furthermore, epistemic peers, though overall similar in intellectual capacity, tend to differ in the exact profile of virtues they possess. Consequently, even assessing exactly the same evidence and arguments, convergence or divergence with one’s peers should still be epistemically relevant if the evidence and arguments are complicated enough that their thorough scrutiny challenges the upper range of human capacity across several intellectual virtues – a condition that the metaphysics of mind appears to meet. Some philosophers are more careful readers of opponents’ views, some more facile with complicated formal arguments, some more imaginative in constructing hypothetical scenarios, etc., and world-class intellectual virtue in any one of these respects can substantially improve the quality of one’s assessments of arguments in the metaphysics of mind. Every philosopher’s preferred metaphysical position is rejected by a substantial proportion of philosophers who are overall approximately as well informed and intellectually virtuous as she is, and who are also in some respects better informed and more intellectually virtuous than she is. Under these conditions, Kelly’s reasons for disregarding peer dissent do not apply, and a high degree of confidence in one’s position is epistemically unwarranted.

Adam Elga (2007) has argued that you can discount peer disagreement if you reasonably regard the fact that the seeming-peer disagrees with you as evidence that, at least on that one narrow topic, that person is not in fact a full epistemic equal. Thus, a materialist might see anti-materialist philosophers of mind, simply by the virtue of their anti-materialism, as evincing less than a perfect level-headedness about the facts. This is not, I think, entirely unreasonable. But it's also fully consistent with still giving the fact of disagreement some weight as a source of doubt. And since your best philosophical opponents will exceed you in some of their intellectual virtues and know some facts and arguments, which they consider relevant or even decisive, which you have not fully considered, you ought to give the fact of dissent quite substantial weight as a source of doubt.

Imagine an array of experts betting on a horse race: Some have seen some pieces of the horses’ behavior in the hours before the race, some have seen other pieces; some know some things about the horses’ performance in previous races, some know other things; some have a better eye for a horse’s mood, some have a better sense of the jockeys. You see Horse A as the most likely winner. If you learn that other experts with different, partly overlapping evidence and skill sets also favor Horse A, that should strengthen your confidence; if you learn that a substantial portion of those other experts favor B or C instead, that should lessen your confidence. This is so even if you don’t see all the experts quite as peers, and even if you treat an expert’s preference for B or C as grounds to wonder about her good judgment.

Try this thought experiment. You are shut in a seminar room, required to defend your favorite metaphysics of mind for six hours (or six days, if you prefer) against the objections of Ned Block, David Chalmers, Daniel Dennett, and Saul Kripke. Just in case we aren’t now living in the golden age of metaphysics of mind, let’s add Kant, Leibniz, Hume, Zhu Xi, and Aristotle too. (First we’ll catch them up on recent developments.) If you don’t imagine yourself emerging triumphant, then you might want to acknowledge that the grounds for your favorite position might not really be very compelling.

It is entirely possible to combine appropriate intellectual modesty with enthusiasm for a preferred view. Consider everyone’s favorite philosophy student: She vigorously champions her opinions, while at the same time being intellectually open and acknowledging the doubt that appropriately flows from her awareness that others think otherwise, despite those others being in some ways better informed and more capable than she is. Even the best professional philosophers still are such students, or should aspire to be, only in a larger classroom. So pick a favorite view! Distribute one’s credences differentially among the options. Suspect the most awesome philosophers of poor metaphysical judgment. But also: Acknowledge that you don't really know.

[For more on disagreement in philosophy see here and here. This post is adapted from my paper in draft The Crazyist Metaphysics of Mind.]

Striking Confirmation of the Spelunker Illusion

In 2010, I worked up a post on what I dubbed The Spelunker Illusion (see also the last endnote of my 2011 book). Now, hot off the press at Psychological Science, Kevin Dieter and colleagues offer empirical confirmation.

The Spelunker Illusion, well-known among cave explorers, is this: In absolute darkness, you wave your hand before your eyes. Many people report seeing the motion of the hand, despite the absolute darkness. If a friend waves her hand in front of your face, you don't see it.

I see three possible explanations:

(1.) The brain's motor output and your own proprioceptive input create hints of visual experience of hand motion.

(2.) Since you know you are moving your hand, you interpret low-level sensory noise in conformity with your knowledge that your hand is in such-and-such a place, moving in such-and-such a way, much as you might see a meaningful shape in a random splash of line segments.

(3.) There is no visual experience of motion at all, but you mistakenly think there is such experience because you expect there to be. (Yes, I think you can be radically wrong about your own stream of sensory experience.)

Dieter and colleagues had participants wave their hands in front of their faces while blindfolded. About a third reported seeing motion. (None reported seeing motion when the experimenter waved his hand before the participants.) Dieter and colleagues add two interesting twists: One is that they add a condition in which participants wave a cardboard silhouette of a hand rather than the hand itself. Under these conditions the effect remains, almost as strong as when the hand itself is waved. The other twist is that they track participants' eye movements.

Eye movements tend to be jerky, jumping around the scene. One exception to this, however, is smooth pursuit, when one stabilizes ones gaze on a moving object. This is not under voluntary control: Without an object to track, most people cannot move their eyes smoothly even if they try. In 1997, Katsumi Watanabe and Shinsuke Shimojo found that although people had trouble smoothly moving their eyes in total darkness, they could do so if they were trying to track their ("invisible") hand motion in darkness. Dieter and colleagues confirmed smooth hand-tracking in blindfolded participants and, strikingly, found that participants who reported sensations of hand motion were able to move their eyes much more smoothly than those who reported no sensations of motion.

I'm a big fan of corroborating subjective reports about consciousness with behavioral measures that are difficult to fake, so I love this eye-tracking measure. I believe that it speaks pretty clearly against hypothesis (3) above.

Dieter and colleagues embrace hypothesis (1): Participants have actual visual experience of their hands, caused by some combination of proprioceptive inputs and efferent copies of their motor outputs. However, it's not clear to me that we should exclude hypothesis (2). And (1) and (2) are, I think, different. People's experience in darkness is not merely blank or pure black, but contains a certain amount (perhaps a lot) of noise. Hypothesis (2) is that the effect arises "top down", as it were, from one's high-level knowledge of the position of one's hand. This top-down knowledge then allows you to experience that noisy buzz as containing motion -- perhaps changing the buzz itself, or perhaps not. (As long as one can find a few pieces of motion in the noise to string together, one might even fairly smoothly track that motion with one's eyes.)

Here's one way to start to pull (1) apart from (2): Have someone else move your hand in front of your face, so that your hand motion is passive. Although this won't eliminate proprioceptive knowledge of one's hand position, it should eliminate the cues from motor output. If efferent copies of motor output drive the Spelunker Illusion, then the Spelunker Illusion should disappear in this condition.

Another possibility: Familiarize participants with a swinging pendulum synchronized with a sound, then suddenly darken the room. If hypothesis (2) is correct and the sound is suggestive enough of the pendulum's exact position, perhaps participants will report still visually experiencing that motion.

Update, April 28, 2014:

Leonard Rosgole and Miguel Roig point out to me that these phenomena were reported in the psychological literature in Hofstetter 1970, Brosgole and Neylon 1973, Brosgole and Roig 1983. If you're aware of earlier sources, I'd be curious to know.