Does the S&M Robot Feel Guilty?

Philosophers of mind typically distinguish experiential states or subjective experiences (like seeing red, feeling pain, or guilt) from intentional states (like believing or wanting) on the basis of their purportedly obvious phenomenal character.  But if subjective experiences really are distinctive because they possess an obvious and unmistakable phenomenal character—in other words, because there is “something it is like to occupy them”—then presumably philosophers and non-philosophers will categorize the same mental states as subjective experiences. Specifically, if philosophers have identified a certain type of mental state as a paradigmatic subjective experience, then we should expect ordinary people to identify the same type of mental state as a subjective experience as well.  But do they?

One method experimental philosophers and cognitive scientists have been using to get at this question of how people categorize mental states draws from philosophical thought experiments, which have long asked us to consider what mental states we would be willing to attribute to other entities. The method assumes that if ordinary people categorize a mental state as an experience insofar as it possesses an unavoidable phenomenal character, then we should expect this to be reflected in their attributions of phenomenal states to other entities. As Sytsma and Machery (2010) write, “we should expect the folk to deny that an entity, be it a simple organism, a simple robot, or a zombie, that lacks phenomenal consciousness can see red just as readily as they deny that it can be in pain” (302).

In a new paper, Sytsma & Machery (S&M) examine how philosophers and the folk attribute different experiential states to a non-human entity: a simple robot named Jimmy [the “S&M robot”] in just this way. In each of their studies, S&M present experimental participants with vignettes in which the S&M robot differentiates boxes on the basis of visual or olfactory cues, receives a high-voltage shock, or meets with a violent, competitor bot. For each scenario, they ask participants whether the robot (e.g.) saw red, smelled bananas, felt pain, or felt angry. What S&M discover is that participants are sometimes willing to ascribe subjective experiences to a robot, but that these ascriptions vary both within and across modalities (they are willing to attribute seeing red and smelling isoamyl acetate to the robot, but not feeling pain or smelling vomit).  In the paper, as Justin Sytsma has recently, succinctly stated, S&M claim that (i) non-philosophers do not distinguish subjective experiences by their common possession of a manifest phenomenal character, and (ii) that perhaps the folk recognize an orthogonal distinction, between mental states that are valenced—such as feeling pain, smelling vomit, or experiencing anxiety—and states which are not valenced—such as seeing red and having a belief.

We greatly admire the work of S&M.   And, we think that S&M have indeed found strong evidence that philosophers and non-philosophers may have different concepts of subjective experience.  However we question their hypothesis that the ordinary concept can be explained in terms of valence.  Recently Mark Phelan and I have been conducting some experiments suggesting that people are perfectly willing to attribute such paradigmatically affective states as smelling vomit and feeling guilt to a robot, so long as these are appropriately related to the function the robot was designed to perform.

What’s more, our experiments demonstrate that the effect is not explained in virtue of the possibility that people assume a robot with a distinctive function to be more complex (we independently varied complexity and function in our studies, and found the same asymmetry for function for a relatively simple as well as a relatively complex robot).  Neither were we able to detect the predicted relationship between state attribution and positive or negative judgments people make regarding state valence.  On the basis of our results, we go on to argue that S&M’s original findings are explained because their test vignettes lead participants to draw specific conclusions about the function of the robots discussed in their studies, and not because the folk recognize a discontinuity between valenced and non-valenced states.

Read more about these findings in a preliminary write-up of our results here. Comments very welcome!

Comments

[Justin Sytsma]

Hey Wes,

The S&M robot? Kinky! And very interesting results. A few thoughts…

First, I don’t think that the paper accurately captures the argument we were making in our paper. As you note in the blog post, and as stated in the opening paragraph of our paper, for example, we had two distinct goals: (1) to present evidence that the folk do not group mental states in the way that philosophers do – specifically, they do not group experiential states together and distinguish them from intentional states (or, in shorthand, that they do not share the philosophical concept of phenomenal consciousness); and, (2) to explore how the folk do group mental states together (suggesting that they distinguish between mental states that essentially have a valence, mental states that do not have a valence, and mental states that have a valence as well as a prominent perceptual component). While the evidence we provide for (2) supports (1), the primary evidence for (1) is from our first study (showing that people treat two paradigmatic examples of phenomenally conscious mental states, seeing red and feeling pain, differently. Having shown that, we then further investigate why they do so. Parts of your paper give a different impression, however.

You note three “mutually exclusive positions” in the literature:

The first position holds that ordinary people group experiential states together and distinguish them from intentional states in roughly the same way that philosophers traditionally have. Alternatively, a second position maintains that, contra philosophical tradition, people do not draw important distinctions between experiential and intentional states. Lastly, the third view, is that people do not ordinary group experiential states together in the way that philosophers have suggested. (p. 2-3)

I’m a little confused by this division of the positions, I suspect because there are two dimensions at play here: (a) do people distinguish experiential states from intentional states (i.e., do people share the philosophical concept of phenomenal consciousness); (b) do people group mental states together in some other way. We argue against (a), which would seem to set us against the first position and in line with the second – that is, we argue in (1) that “people do not draw important distinctions between experiential and intentional states.” (Side note: I didn’t think that Arico et al, 2011, had argued against (a), as I understand it.)

You go on to write: “S&M in turn contend that this valenced conception of subjective experience raises problems for a philosophical tradition which emphasizes the manifest phenomenal character of all subjective experiences.” (p. 3). That doesn’t seem quite right, however, in that this claim does not depend on the valence hypothesis (although the evidence we present for the valence hypothesis lends further support for the claim). That is, the point of (1) is to argue that people do not group phenomenally conscious mental states together, and this is distinct from (2).

You continue: “We argue from our own experimental data that, contra S&M, people do not ordinarily differentiate experiences according to whether they are valenced. This undercuts S&M’s criticism of philosophical tradition, but our discussion does not thereby support that tradition.” (p. 3). But, this would only undercut the secondary support for the criticism of that we draw from (2). We present evidence that the folk do not group mental states together in the way that philosophers traditionally do that is independent of the positive claim for the valence hypothesis. (You do go on to essentially distinguish between (1) and (2) in Section 1, but this seems to me to be somewhat at odds with the characterization in the introduction.)

Second, I think it is important to note that the valence hypothesis does not simply say that people distinguish between valenced and non-valenced mental states, but instead draws a three-fold distinction as noted above. For example, we write that “people distinguish between states that essentially have a valence (such as feeling pain or anger), those that do not have a valence (such as seeing red), and those that have a valence as well as a prominent perceptual component (such as smelling banana).” (+2). This is important to note, in part, because at least some of the mental states you test are those that we explicitly place into the third grouping. Thus, we write that:

“For a perceptual state like smelling banana that has an associated valence, however, the folk are divided. We conjecture that they are divided because they hold that while Jimmy is capable of perceiving the scent of banana, he is incapable of having the valence; he is incapable of liking the scent of banana. The folk are divided, here, because unlike pain where its valence is essential to it being pain, for smell they see the valence as potentially disassociating from the perceptual aspect.” (+17)

We go on to say that “if this explanation is correct, then we would expect that downplaying the valence reading for smell will increase the folk’s willingness to attribute the state to the robot Jimmy.” (+17). One plausible explanation of the results of your first study, then, is that in specifying the function of the robot, you downplay the valence reading for odors that are relevant to that function. And, what we see in Figure 2, for example, is that people were mixed on whether smoothie-function-Jimmy smelled vomit, but agreed that waste-function-Jimmy smelled vomit.

It is worth noting, here, that I don’t take us to be committed to denying that other factors besides valence play a role in folk attributions of mental states. I think your results clearly indicate that function plays an important role (although see below about function-specific complexity). But, I’m less convinced that they show that valence does not play a role. In fact, it seems to me that taken together your experiments suggest that valence does matter! Thus, we see that people were mixed about whether smoothie-function-Jimmy smells vomit, while tending to deny that lifting-function-Jimmy felt guilt.

Finally, I’m a little unsure about the controls you run for complexity. The reason is that complexity is complex. Specifically, it seems like we should distinguish between general complexity and function-specific complexity. It strikes me that while you test general complexity as a variable, this does not really control for complexity. That is, it still seems plausible that regardless of what is said about Jimmy’s general complexity, in specifying a function the implication is that he is more complex with regard to that function than when no function is specified. For examples, it strikes me as reasonable to think that simple smoothie-function-Jimmy is more complex with regard to detecting some odors (such as banana) than is simple waste-function-Jimmy, and likewise that complex smoothie-function-Jimmy is more complex with regard to detecting some odors (such as banana) than is complex waste-function-Jimmy. Given that it looks like there was an effect for general complexity, I am unsure how varying the general complexity of Jimmy acts as a control.

[Wesley Buckwalter]

Hey Justin, thanks for the comments. Glad you like the title (Josh Knobe can’t stop giggling when he hears it).

These comments are really helpful, I think you are totally right that we could have described the dialectic slightly more carefully. First, let me start off by saying that we never intended to question (1) above, only (2). Maybe this is more clear in the actual post than the paper itself where we say “we think that S&M have indeed found strong evidence that philosophers and non-philosophers may have different concepts of subjective experience. However we question their hypothesis that the ordinary concept can be explained in terms of valence.” So we take our paper not as questioning that people distinguish experiential states differently, but rather just S&M’s positive thesis that (again, that if they are different) this difference is best not explained by the valence hypothesis offered.

With this clarified, maybe it will help with some of the questions you raise. So for instance you say “You (we) go on to write: “S&M in turn contend that this valenced conception of subjective experience raises problems for a philosophical tradition which emphasizes the manifest phenomenal character of all subjective experiences.” (p. 3). That doesn’t seem quite right, however, in that this claim does not depend on the valence hypothesis (although the evidence we present for the valence hypothesis lends further support for the claim).” So, I don’t think we ever claim that the problems raised for philosophical tradition themselves depend on your valence hypothesis directly (nonetheless it is true that the valence hypothesis would if true). Alternatively, we think that while there may be problems based on your data (as you mention for instance in your first experiment) the game is afoot to provide the best explanation of those data re coming up with a theory that characterizes the ordinary concept of subjective experience.

Now, about those data. The key point of the B&P studies is that there's an alternative explanation of the S&M findings, namely that function is important. That is could people’s state ascriptions in S&M not (or not only) speak to folk beliefs about the nature of experiential states, but rather to folk beliefs about the function of the entity to which those experiential states are attributed? By your own admission, our studies provide precisely this kind of evidence. Of course, this impact of function assumptions on people’s state judgments might be compatible with the valence hypothesis insofar as there can be additional factors besides valence playing a role in folk attributions of mental states (as I would hope!). However, our claim is a bit more ambitious in the paper. We offer function not only as a thing important to people’s judgments, but also as a potential alternative explanation to S&M data.

So about that alternative explanation, does it trouble you that there has been no evidence (and in our studies especially) of any kind of direct relationship between people’s actual valence judgments and people’s state attributions? For instance, I was pretty shocked in our studies that not only could we not find a moderating relationship here…we could not even detect simple correlations between sig differences in people’s affect judgments and sig differences in state attributions (even though to do so would on the surface reflect a boring statistical fact given these differences). But if we’re not finding relationships between people’s valence judgments (in the paper we measure this on what we call an “affect scale” of positive and negative valence) then why not suspect something else about the states in questions that is doing the work here in distinguishing states? Our suggestion is that it could be the way they are hooking up to the function of the entity to which they are ascribed.

[Justin Sytsma]

Hey Wes (and Mark!– I should have used “you two” for this, or something, but that seems awkward),

I definitely think what you two are up to is clearer in the blog post than the paper—and think that the paper would benefit from some clarification along that score. (I expect that I would find it helpful, in part, because I am far more invested in the first claim—that the folk do not have share the philosophers’ concept of phenomenal consciousness—than the valence hypothesis, having argued for the first claim in several other papers, as well as an important chapter of my dissertation!)

With regard to the studies, I agree that your results suggest against the valence hypothesis, although with a few reservations. One reservation I have is that I’m not convinced by the control you did for complexity. I’m thinking a better approach, with regard to testing the valence hypothesis specifically, would be to specify a general function for the robot in each case, not different functions that would seem to imply that the robot is more complex with regard to certain perceptual tasks then others. (I’ve thrown up a few probes to look at this and will let you know what I find.)

Another reservation is that from what I can tell from your data, it looks like you found a difference between off-function olfaction and off-function guilt, which would seem to fit with the valence hypothesis. From the other direction, it doesn’t seem like this is well-explained in terms of the function hypothesis. What do you make of mental state ratings being higher for smoothie-function-Jimmy for vomit than for lifting-function-Jimmy for guilt?

A third reservation relates to your point about actual valence judgments (which is well taken). If I’m following the paper correctly, you asked about the valence of the odors used in your first study and not the second. (Is that correct? Did you ask the valence question after the Jimmy probe?) But, in our paper we argue that people are mixed about olfaction – it has a valence as well as a prominent perceptual component. And we argue that focusing attention on the perceptual task / downplaying the valence would increase attributions of smell to Jimmy. As such, valence ratings would be expected to pull away from the mental state ratings as those ratings go up. I couldn’t tell from the stats you report, though, whether there was any correlation for smoothie-function-Jimmy for Vomit, or other cases with ratings close to 4. Of course, we would also expect that valence ratings would correlate with how likely people are to focus on valence in attributing the corresponding mental state to Jimmy. (But, I didn’t take us to be committed to a direct correlation; rather, we argue that most people do not find that a simple robot is capable of having mental states with an associated valence, whether strong or weak.)

Let me hasten to emphasize, though, that despite my reservations, I think this is very cool work! And I’m very happy to see this side of the debate get explored further!!

[Mark Phelan]

Thanks for the careful comments, Justin. They are very helpful to us in reformulating what—I hasten to point out—is a very preliminary discussion of our results. I’ll probably be picking up most of the discussion from now on, since Wes is on the verge of nuptial bliss.

I think it’s best to leave the exegetical discussions aside here (though your points are well taken). Nonetheless, I have a few quick responses to other parts of your comments:

The first concerns the issue of our testing states that belong in the third category, that is, states that have a valence as well as a “prominent perceptual component”. First of all, I’m not sure what a prominent perceptual component is such that smelling a banana has one and seeing red or feeling pain does not. It might be helpful just to explicate this. But, perhaps more importantly, I’m not sure why specifying the function of the robot in our first study would, “downplay the valence reading for odors that are relevant to that function” (emphasis mine). Why wouldn’t thinking of the function of making smothies be just as likely to remind our participants of their own past smoothie making experiences, and thereby to remind them of the valence of banana smells and etc? Could we operationalize and devise a test for downplaying (or up-playing) valence?

My second comment bears on complexity. I like the distinction between general complexity (g complexity) and function-specific complexity (fs complexity). I think it’s important to note though, that g complexity and fs complexity interact in dynamic ways. Participants in our studies are asked about 1 of 2 robots that differ in general complexity: A robot designed at a State University of non-descript components, and a robot designed in the Ivies with top-of-the-line components. As you rightly note, in addition to the different levels of g complexity, a distinctive complexity may come along with function, which is a variable we also manipulate. When the robot is designed to make smoothies, dispose of biological waste, or be a friend to the elderly, participants likely suppose a competent designer who saw to it the robot had all the components relevant to its specific job. But if the designer could also avail himself of top-of-the-line components, as the Ivy-bot’s designer could, then the fs complexity would be even more complex. So, the State U bots will differ in fs complexity relative to their function—the robot designed to be a friend to the elderly will be more fs complex for the relevant functions than the robot designed to lift things. But there will be a greater disparity for the Ivy-league bots. The Ivy-bot designed to be a friend to the elderly will be even more fs complex for the relevant functions than his Ivy-bot counterpart, because the components which enable him to perform the relevant functions are just generally very complex. Thus, if greater complexity is driving our effect, we would expect a greater difference between state attributions for the robots in the Ivy conditions than in the State conditions, which we don’t find. In fact, the differences between mean responses for these categories is exactly the same.

Of course, that’s all a bit speculative. Our study wasn’t designed to find such differences, and it’s just a post-hoc consideration given the findings. A more concrete point is this: Obviously, specifying that something has a function leads us to attribute more complexity relative to that function, since we suppose the thing has components relevant to performing that function. But we certainly wouldn’t want to conclude that a simple robot couldn’t occupy a particular mental state because the state is valenced when the robot in fact just lacks components necessary to occupy the state. I don’t think my remote control gets depressed when I turn on the Real Housewives of L.A., but I’m fairly certain that’s not importantly due to the valence of depression, rather it’s due to the structure of my remote. The worry we’re raising about your studies is akin to this: Perhaps your studies aren’t getting at differences in how people think about mental states. Rather they may be due to differences in how people think about the robots in your studies. I think the differences you found between seeing red and smelling banana on the one hand, versus feeling anger on the other, is very telling in this respect. Each of your robots is specified to have, “a scent detector, video camera for eyes, wheels for moving about, and two grasping arms with touch sensors that he can move objects with.” This includes machinery relevant to seeing red and smelling bananas. But there’s no mention here of any machinery relevant to feeling anger—whatever that machinery may be. Thus, it’s perfectly reasonable to suppose that the relatively low rating for anger is just due to a failure to conceive of the robot as having the relevant machinery. In this light, our study corrects for this problem, by leading participants to think of a designer who gave a robot machinery relevant to its specific task, whatever that machinery is. And, indeed, once that machinery is onboard, even a simple robot is thought capable of occupying valenced states. Thus, our studies present a problem for interpreting the evidence supposed to support the valence hypothesis.

Anyway, these are dicey issues and I think we’re going to have to exercise a lot of care in operationalizing terms in this neighborhood and thinking about what studies show and how to disambiguate various hypotheses. I’m excited to see the studies you’ve thrown up, and whatever results they deliver! And I look forward to talking to you and others about this here and elsewhere in the future.

[Justin Sytsma]

Congrats Wes! That is wonderful news!!

Mark, I'm glad my comments have been helpful. And definitely an interesting discussion! More thoughts tomorrow (right now I'm too exhausted from a long week to think well).

[Justin Sytsma]

Hey Mark,

Again, let me say that this is very interesting work. And, I think you are right to push on the valence hypothesis. In fact, Edouard and I have been doing some further research along these lines, conducting a two-stage study in which we got valence ratings for a range of odors first, then testing a range of those odors with the Jimmy probes. Although this is still in process, the results aren’t looking promising for the valence hypothesis as formulated. It could be that a different positive account of the differences in folk attributions is needed. I’m not sure that the function hypothesis does the trick, however. In large part this is because I’m not sure how it would account for the divergence between seeing red and feeling pain (the asymmetry that led us toward the valence hypothesis) although, certainly, further testing might reveal that it does; and, in part, because it still doesn’t seem to account for the divergence you found between odors and guilt (for the simple off-function robot, for example). From the current data, it seems to me that multiple factors are probably involved. (It might be that, following Knobe, embodiment is one factor. With regard to pain, this would fit with other work of mine indicating the folk are naïve realists both about qualities like colors and about pains.)

With regard to your points about your studies: In putting forward the three-fold distinction mentioned, we weren’t denying that seeing red / feeling pain have a prominent perceptual component, rather the key points are whether they have a valence and whether that valence is thought to be essential to having the mental state. With regard to seeing red, the claim is that for the cases at issue, at least, it isn’t thought to have a valence. Feeling pain, on the other hand, is thought to essentially have a valence. Likewise, many odors are thought to have a valence, but we claim that that valence isn’t thought to be essential to the state: That is, we claim that there is both a perceptual-reading and a valence-reading for states like smelling vomit. On the perceptual-reading we would expect people to ascribe smelling vomit to Jimmy, while on the valence-reading we would expect that they wouldn’t. With regard to why specifying the function of the robot might downplay the valence-reading for odors that are relevant to that function, the reason I had in mind is that the function you specify implies that they would be able to detect those odors. This would seem to me to emphasize the perceptual-reading for those states, and thereby downplay the valence-reading. I’m not sure how to operationalize this, but this fits in with my concerns about function-specific complexity. It strikes me that insofar as part of the argument against the valence hypothesis is that in our original probes function was unspecified, the test I suggested might be a good one – specifying a general function for all of the cases that isn’t specific to certain odors.

With regard to general complexity and function-specific complexity, that is a good point. I’m not sure, however, that I would have expected function-specific complexity to interact with general complexity in that way. I guess that it seems plausible to me that people would treat these independently. Certainly, I’m not sure how large of an effect should be expected, if one is expected at all, so I’m not sure whether your studies would detect the effect.

“A more concrete point is this: Obviously, specifying that something has a function leads us to attribute more complexity relative to that function, since we suppose the thing has components relevant to performing that function. But we certainly wouldn’t want to conclude that a simple robot couldn’t occupy a particular mental state because the state is valenced when the robot in fact just lacks components necessary to occupy the state.”

It seems like the comparison between types of odors controls for this (although you are right that this raises a difficulty for ascriptions across modality, and your point about anger is well taken). The problem that function-specific complexity raises, however, is that it isn’t being held constant across the different odors. And insofar as complexity is expected to make a difference, this raises a confound with regard to testing the valence hypothesis.

[John Jones]

I wonder how far the experiment itself falls victim to the discourse of popular animism. For example, the way the S&M object is introduced to subjects would direct any subsequent discourse about it. Some thing that is introduced as being able to differentiate, receive, accept, "meet" or as "Jimmy" will signal to participants that they have permission to talk about it in the philosphically uncommitted banter of animism.

There is also a complication in that "mental states" that may be attributed to a robot may not always be attributed to other animistic anthropomorphisms such as dolls (and vice versa).

Finally, what would it mean for a person to "attribute" a mental state or subjectivity "to" a thing like a robot (or doll)? The attribution, I am sure, isn't like that made by the scientist/biologist. More likely it isn't an attribution at all, even less one of a mental "state". Similar considerations can be made in respect of "complexity". There is no link in animist discourse between function and complexity, though there may be a link for scientist and philosopher. More needs to be examined about this supposed meeting of science and animism.

[Mark Phelan]

Sorry for my delay in responding, Justin…it’s been a busy week.
First, thanks for your candor about the valence hypothesis. One thing I like about working in experimental philosophy is the intellectual honesty of many of the researchers, yourself included.

As for the point about how function could deal with the asymmetry in responses you found between seeing red and feeling pain, I was thinking that the particular result might be explained in the same way I was suggesting the differences between smelling x and feeling anger were explained—that is, the way Jimmy is specified includes a description of machinery relevant to seeing red but none for feeling pain. (I know he has “touch sensors”, but I’m suggesting touching and feeling pain are different capacities with distinctive functions and machinery.)

There’s a more general point here about pain, though, that you might be making. It is after all notoriously difficult to give a teleofunctional account of pain. You might think then that the function hypothesis is vacuous when it comes to pain…it can’t say people will think S is in pain if S is designed to perform the behaviors pain experience makes possible, because there is no common understanding of what those behaviors are. That strikes me as a reasonable point—and also as a reason to avoid using pain as our paradigm of a phenomenal state!

Certainly, I agree with you that multiple factors are involved in phenomenal state attribution. (Who could think otherwise?) And, “specifying a general function for all of the cases that isn’t specific to certain odors,” also seems like a good way to proceed in investigating these issues. Though it would certainly be difficult to come up with such a function for all the relevant states. (Maybe we can brainstorm about this more off the blog?)

John:

I am very sensitive to your concerns about the difference between what is said and what is meant. (See my paper, “Thinking Things and Feeling Things,” with Adam Arico and Shaun Nichols.) I think that the best way to study these issues is by investigating how people paraphrase different verbal ascriptions, and I’d like to devise some experiments to look into this for robots.

[Josh Weisberg]

Mark and Wesley--very cool paper, and a nice discussion here. And Justin, I'd be interested in your answer as well:

I wonder whether the issue here isn't one about the external nature of perception vs. the internal nature of feeling. Let me explain. Red is (naively?) thought of as a thing in the world--the red is "on" the apple. Pain, on the other hand, is in me. What's more, as the debate over representationalism in consciousness studies (see Tye, etc.) seems to show, what we seem to be aware of when we're aware of pain is not bodily damage but the feeling of pain itself (or anyway, there's something different about pain and seeing red along these lines). Interestingly, people seem to vacillate (no study--just intuition from discussions with philosophers and students!) about whether we're aware of the banana when we smell it or the smell of the banana. Taste is even more "internal." And shame and anger seem to be like pain--I'm not aware of what my anger is about (at least not directly)--I'm aware of the feeling of anger.

If you follow this rather impressionistic taxonomy, it may explain what folks are willing to attribute to robots. Robots can detect things in the world just fine--even red things or bananas. But if the probes are interpreted as asking about the "inner feelings" of robots, people balk. To the extent that mental states are "transparent" robots can have them--we can think of those states in terms of what they pick out in the world. But when it comes to states like pain and anger, there is not an obvious object in the world to tag them to. So they must be inner feelings, things that robots cannot have (why people might think this remains an interesting question). Smells and tastes (and maybe sounds) are mixed states--sometimes we say we smell the banana (something a robot could do--detecting bananas by chemotaxis). But sometimes we say we are aware of a banana taste--the taste seems to be in me, rather than "on" the banana.

Do y'all control for this? Does this sound plausible? This may (sadly, from my perspective) vindicate the philosophers' concept of phenomenal consciousness. One key question is can we find a way to probe subjects about reddish experiences, rather than the red thing detected? Perhaps Jimmy the robot and Timmy the human both detect (see, even experience) red, but only Timmy has a reddish experience. Maybe a relevant question is whether robots can imagine red? Then there is no red present in the world. But perhaps imagination piggy-backs on perception too closely.

Anyway, cool stuff!

[Justin Sytsma]

Hey Josh,

In our original paper that was the external/internal explanation was the first we tested. But, based on the results for smell, we tentatively concluded that it didn't explain the pattern of results.

The new results that have been coming in on smell, however, are higher than we had gotten. And, I think you are right -- this suggests that the external/internal hypothesis should be revisited. In fact, in some other work (http://philsci-archive.pitt.edu/5141/), I've given evidence indicating that the folk tend to be naive realists about both colors and pains, which lends itself to the external/internal hypothesis.