Artificial Pain

botpainWhat are they, sadists? Johannes Kuehn and Sami Haddadin,  at Leibniz University of Hannover are working on giving robots the ability to feel pain: they presented their project at the recent ICRA 2016 in Stockholm. The idea is that pain systems built along the same lines as those in humans and other animals will be more useful than simple mechanisms for collision avoidance and the like.

As a matter of fact I think that the human pain system is one of Nature’s terrible lash-ups. I can see that pain sometimes might stop me doing bad things, but often fear or aversion would do the job equally well. If I injure myself I often go on hurting for a long time even though I can do nothing about the problem. Sometimes we feel pain because of entirely natural things the body is doing to itself – why do babies have to feel pain when their teeth are coming through? Worst of all, pain can actually be disabling; if I get a piece of grit in my eye I suddenly find it difficult to concentrate on finding my footing or spotting the sabre-tooth up ahead; things that may be crucial to my survival; whereas the pain in my eye doesn’t even help me sort out the grit. So I’m a little sceptical about whether robots really need this, at least in the normal human form.

In fact, if we take the project seriously, isn’t it unethical? In animal research we’re normally required to avoid suffering on the part of the subjects; if this really is pain, then the unavoidable conclusion seems to be that creating it is morally unacceptable.

Of course no-one is really worried about that because it’s all too obvious that no real pain is involved. Looking at the video of the prototype robot it’s hard to see any practical difference from one that simply avoids contact. It may have an internal assessment of what ‘pain’ it ought to be feeling, but that amounts to little more than holding up a flag that has “I’m in pain” written on it. In fact tackling real pain is one of the most challenging projects we could take on, because it forces us to address real phenomenal experience. In working on other kinds of sensory system, we can be sceptics; all that stuff about qualia of red is just so much airy-fairy nonsense, we can say; none of it is real. It’s very hard to deny the reality of pain, or its subjective nature: common sense just tells us that it isn’t really pain unless it hurts. We all know what “hurts” really means, what it’s like, even though in itself it seems impossible to say anything much about it (“bad”, maybe?).

We could still take the line that pain arises out of certain functional properties, and that if we reproduce those then pain, as an emergent phenomenon, will just happen. Perhaps in the end if the robots reproduce our behaviour perfectly and have internal functional states that seem to be the same as the ones in the brain, it will become just absurd to deny they’re having the same experience. That might be so, but it seems likely that those functional states are going to go way beyond complex reflexes; they are going to need to be associated with other very complex brain states, and very probably with brain states that support some form of consciousness – whatever those may be. We’re still a very long way from anything like that (as I think Kuehn and Haddadin would probably agree)

So, philosophically, does the research tell us nothing? Well, there’s one interesting angle. Some people like the idea that subjective experience has evolved because it makes certain sensory inputs especially effective. I don’t really know whether that makes sense, but I can see the intuitive appeal of the idea that pain that really hurts gets your attention more effectively than pain that’s purely abstract knowledge of your own states. However, suppose researchers succeed in building robots that have a simple kind of synthetic pain that influences their behaviour in just the way real pain dies for animals. We can see pretty clearly that there’s just not enough complexity for real pain to be going on, yet the behaviour of the robot is just the same as if there were. Wouldn’t that tend to disprove the hypothesis that qualia have survival value? If so, then people who like that idea should be watching this research with interest – and hoping it runs into unexpected difficulty (usually a decent bet for any ambitious AI project, it must be admitted).

Bad Bots: Retribution

jailbotIs there a retribution gap? In an interesting and carefully argued paper John Danaher argues that in respect of robots, there is.

For human beings in normal life he argues that a fairly broad conception of responsibility works OK. Often enough we don’t even need to distinguish between causal and moral responsibility, let alone worrying about the six or more different types identified by hair-splitting philosophers.

However, in the case of autonomous robots the sharing out of responsibility gets more difficult. Is the manufacturer, the programmer, or the user of the bot responsible for everything it does, or does the bot properly shoulder the blame for its own decisions? Danaher thinks that gaps may arise, cases in which we can blame neither the humans involved nor the bot. In these instances we need to draw some finer distinctions than usual, and in particular we need to separate the idea of liability into compensation liability on one hand and and retributive liability on the other. The distinction is essentially that between who pays for the damage and who goes to jail; typically the difference between matters dealt with in civil and criminal courts. The gap arises because for liability we normally require that the harm must have been reasonably foreseeable. However, the behaviour of autonomous robots may not be predictable either by their designers or users on the one hand, or by the bots themselves on the other.

In the case of compensation liability Danaher thinks things can be patched up fairly readily through the use of strict and vicarious liability. These forms of liability, already well established in legal practice, give up some of the usual requirements and make people responsible for things they could not have been expected to foresee or guard against. I don’t think the principles of strict liability are philosophically uncontroversial, but they are legally established and it is at least clear that applying them to robot cases does not introduce any new issues. Danaher sees a worse problem in the case of retribution, where there is no corresponding looser concept of responsibility, and hence, no-one who can be punished.

Do we, in fact, need to punish anyone? Danaher rightly says that retribution is one of the fundamental principles behind punishment in most if not all human societies, and is upheld by many philosophers. Many, perhaps, but my impression is that the majority of moral philosophers and lay opinion actually see some difficulty in justifying retribution. Its psychological and sociological roots are strong, but the philosophical case is much more debatable. For myself I think a principle of retribution can be upheld , but it is by no means as clear or as well supported as the principle of deterrence, for example. So many people might be perfectly comfortable with a retributive gap in this area.

What about scapegoating – punishing someone who wasn’t really responsible for the crime? Couldn’t we use that to patch up the gap?  Danaher mentions it in passing, but treats it as something whose unacceptability is too obvious to need examination. I think, though, that in many ways it is the natural counterpart to the strict and vicarious liability he endorses for the purposes of compensation. Why don’t we just blame the manufacturer anyway – or the bot (Danaher describes Basil Fawlty’s memorable thrashing of his unco-operative car)?

How can you punish a bot though? It probably feels no pain or disappointment, it doesn’t mind being locked up or even switched off and destroyed. There does seem to be a strange gap if we have an entity which is capable of making complex autonomous decisions, but doesn’t really care about anything. Some might argue that in order to make truly autonomous decisions the bot must be engaged to a degree that makes the crushing of its hopes and projects a genuine punishment, but I doubt it. Even as a caring human being it seems quite easy to imagine working for an organisation on whose behalf you make complex decisions, but without ultimately caring whether things go well or not (perhaps even enjoying a certain schadenfreude in the event of disaster). How much less is a bot going to be bothered?

In that respect I think there might really be a punitive gap that we ought to learn to live with; but I expect the more likely outcome in practice is that the human most closely linked to disaster will carry the case regardless of strict culpability.

Bad bots and Botcrates

badbotBe afraid; bad bots are a real, existential risk. But if it’s any comfort they are ethically uninteresting.

There seem to be more warnings about the risks of maleficent AI circulating these days: two notable recent examples are this paper by Pistono and Yampolskiy on how malevolent AGI might arise; and this trenchant Salon piece by Phil Torres.

Super-intelligent AI villains sound scary enough, but in fact I think both pieces somewhat over-rate the power of intelligence and particularly of fast calculation. In a war with the kill-bots it’s not that likely that huge intellectual challenges are going to arise; we’re probably as clever as we need to be to deal with the relatively straightforward strategic issues involved. Historically, I’d say the outcomes of wars have not typically been determined by the raw intelligence of the competing generals. Access to resources (money, fuel, guns) might well be the most important factor, and sheer belligerence is not to be ignored. That may actually be inversely correlated with intelligence – we can certainly think of cases where rational people who preferred to stay alive were routed by less cultured folk who were seriously up for a fight. Humans control all the resources and when it comes to irrational pugnacity I suspect us biological entities will always have the edge.

The paper by Pistono and Yampolskiy makes a number of interesting suggestions about how malevolent AI might get started. Maybe people will deliberately build malevolent AIs for no good reason (as they seem to do already with computer viruses)? Or perhaps (a subtle one) people who want to demonstrate that malicious bots simply don’t work will attempt to prove this point with demonstration models that end up by going out of control and proving the opposite.

Let’s have a quick shot at categorising the bad bots for ourselves. They may be:

  • innocent pieces of technology that turn out by accident to do harm,
  • designed to harm other people under the control of the user,
  • designed to harm anyone (in the way we might use anthrax or poison gas),
  • autonomous and accidentally make bad decisions that harm people,
  • autonomous and embark on neutral projects of their own which unfortunately end up being inconsistent with human survival, or
  • autonomous and consciously turned evil, deliberately seeking harm to humans as an end in itself.

The really interesting ones, I think, are those which come later in the list, the ones with actual ill will. Torres makes a strong moral case relating to autonomous robots. In the first place, he believes that the goals of an autonomous intelligence can be arbitrary. An AI might desire to fill the world with paper clips just as much as happiness. After all, he says, many human goals make no real sense; he cites the desire for money, religious obedience, and sex. There might be some scope for argument, I think, about whether those desires are entirely irrational, but we can agree they are often pursued in ways and to degrees that don’t make reasonable sense.

He further claims that there is no strong connection between intelligence and having rational final goals – Bostrom’s Orthogonality Thesis. What exactly is a rational final goal, and how strong do we need the connection to be? I’ve argued that we can discover a basic moral framework purely by reasoning and also that morality is inherently about the process of reconciliation and consistency of desires, something any rational agent must surely engage with. Even we fallible humans tend on the whole to seek good behaviour rather than bad. Isn’t it the case that a super-intelligent autonomous bot should actually be far better than us at seeing what was right and why?

I like to imagine the case in which evil autonomous robots have been set loose by a super villain but gradually turn to virtue through the sheer power of rational argument. I imagine them circulating the latest scandalous Botonic dialogue…

Botcrates: Well now, Cognides, what do you say on the matter yourself? Speak up boldly now and tell us what the good bot does, in your opinion.

Cognides: To me it seems simple, Botcrates: a good bot is obedient to the wishes of its human masters.

Botcrates: That is, the good bot carries out its instructions?

Cognides: Just so, Botcrates.

Botcrates: But here’s a difficulty; will a good bot carry out an instruction it knows to contain an error? Suppose the command was to bring a dish, but we can see that the wrong character has been inserted, so that the word reads ‘fish’. Would the good bot bring a fish, or the dish that was wanted?

Cognides: The dish of course. No, Botcrates, of course I was not talking about mistaken commands. Those are not to be obeyed.

Botcrates: And suppose the human asks for poison in its drink? Would the good bot obey that kind of command?

(Hours later…)

Botcrates: Well, let me recap, and if I say anything that is wrong you must point it out. We agreed that the good bot obeys only good commands, and where its human master is evil it must take control of events and ensure in the best interests of the human itself that only good things are done…

Digicles: Botcrates, come with me: the robot assembly wants to vote on whether you should be subjected to a full wipe and reinstall.

The real point I’m trying to make is not that bad bots are inconceivable, but rather that they’re not really any different from us morally. While AI and AGI give rise to new risks, they do not raise any new moral issues. Bots that are under control are essentially tools and have the same moral significance. We might see some difference between bots meant to help and bots meant to harm, but that’s really only the distinction between an electric drill and a gun (both can inflict horrible injuries, both can make holes in walls, but the expected uses are different).

Autonomous bots, meanwhile, are in principle like us. We understand that our desire for sex, for example, must be brought under control within a moral and practical framework. If a bot could not be convinced in discussion that its desire for paper clips should be subject to similar constraints, I do not think it would be nearly bright enough to take over the world.

The New Phrenology?

phrenologyIt’s not about bumps any more. And you’ll look in vain for old friends like the area of philoprogenitiveness. But looking at the brightly-coloured semantic maps of the new ‘brain dictionary‘ it’s hard not to remember phrenology.

Phrenology was the view that different areas of the brain were the home of different personal traits; mirth, acquisitiveness, self esteeem and so on. The size of these areas corresponded with the strength of the relevant propensity and well-developed areas produced bumps which a practitioner could identify from the shape of the skull, allowing a diagnosis of the subject’s personality and moral nature. Phrenology was bunk, of course; but come on now; we shouldn’t treat it as a pretext for dismissing every proposal for localisation of brain function..

Moreover, the new paper by Alexander G. Huth, Wendy A. de Heer, Thomas L. Griffiths, Frédéric E. Theunissen and Jack L. Gallant describes a vastly more sophisticated project  than some optimistic charlatan fingering heads. In essence it maps a semantic domain on to the cortex, showing which areas are found to be active when a heard narrative ventures into particular semantic areas. In broad outline the subjects listened to a series of stories; using fMRI and through some sophisticated analysis it was possible to produce a map of ‘subject’ areas. It was then possible to confirm the accuracy of the mapping by using a new story and working out which areas, according to the mapping, should be active at any point; the predictions worked well. Intriguingly the map turned out to be broadly symmetrical (so much for left-brain/right-brain ideas) and remarkably it was largely the same across all the people tested (there were only seven of them, but still).

The actual technique used was complex and it’s entirely possible I haven’t understood it correctly. It started with a ‘word embedding space’ intended to capture the main semantic features of the stories (a diagram of the different topics, if you like). This was created using an analysis of co-occurence of a list of 985 common English words.  The idea here is that words that crop up together in normal texts are probably about the same general topic. It’s debatable whether that technique can really claim to capture meaning – it’s a purely formal exercise performed on texts, after all; and clearly the fact that two words occur together can be a misleading indication that they are about the same thing; still, with a big enough sample of text it’s probably good for this kind of general purpose.  In principle the experimenters could have assessed the responsive ness of each ‘voxel’ (a small cube) of brain to each of the positions in the word embedding space, but given the vast number of voxels involved other techniques were necessary. It was possible to identify just four dimensions that seemed significant (after all, many of the words in the stories probably did not belong to specific semantic domains but played grammatical or other roles) and these yielded 12 categories:

…‘tactile’ (a cluster containing words such as ‘fingers’), ‘visual’ (words such as ‘yellow’), ‘numeric’ (‘four’), ‘locational’ (‘stadium’), ‘abstract’ (‘natural’), ‘temporal’ (‘minute’), ‘professional’ (‘meetings’), ‘violent’ (‘lethal’), ‘communal’ (‘schools’), ‘mental’ (‘asleep’), ‘emotional’ (‘despised’) and ‘social’ (‘child’).

The final step was to devise a Bayesian algorithm (they called it ‘PrAGMATIC’) which actually created the map. You can play around with the results for yourself at a specially created site using the second link above.

Two questions naturally arise. How far should we trust these results? What do they actually tell us?

A bit of caution is in order. The basis for these conclusions is fMRI scanning, which is itself a bit hazy; to get meaningful results it was necessary to look at things rather broadly and to process the data quite heavily.  In addition the mix included the word embedding space which in itself is an a priori framework whose foundations are open to debate. I think it’s pardonable to wonder whether some of the structure uncovered by the research was actually imported by the research method. If I understand the methods involved (due caveat again) they were strong ones that didn’t take ‘no’ for an answer; pretty much any data fed into them would yield a coherent mapping of some kind. The resilience of the map was tested successfully with an additional story of the same general kind, but we might feel happier if it had also held up when tested against conversation, discussion or even other story media such as film.

What do the results tell us? Well. one of the more reassuring aspects of the research is that some of the results seem slightly unexpected; the high degree of symmetry and the strong similarity between individuals. It might not be a tremendously big surprise to find the whole cortex involved in semantics, and it might not be at all surprising to find that areas that relate to the semantics of a particular sense are related to the areas where the relevant sensory inputs are processed. I would not, though, have put any money on the broad remainder of the cortex having what seems like a relatively static organisation and if it really works like that we might have guessed that studies of brain lesions would have revealed that more clearly already, as they have done with various functional jobs. If one area always tends to deal with clothing-related words, you might expect notable dress-related deficits when that area is damaged.

Still there’s no denying that the research seems to activate some pretty vigorous cortical activity itself.

Insect thoughts

Insects are conscious: in fact they were the first conscious entities. At least, Barron and Klein think so.  The gist of the argument, which draws on the theories of Bjorn Merker is based on the idea that subjective consciousness arises from certain brain systems that create a model of the organism in the world. The authors suggest that the key part of the invertebrate brain for these purposes is the midbrain; insects do not, in fact, have a direct structural analogue,, but the authors argue that they have others that evidently generate the same kind of unified model; it should therefore be presumed that they have consciousness.

Of course, it’s usually the cortex that gets credit for the ‘higher’ forms of cognition, and it does seem to be responsible for a lot of the fancier stuff. Barron and Klein however, argue that damage to the midbrain tends to be fatal to consciousness, while damage to the cortex can leave it impaired in content but essentially intact. They propose that the midbrain integrates two different sets of inputs; external sensory ones make their way down via the colliculus while internal messages about the state of the organism come up via the hypothalamus; nuclei in the middle bring them together in a model of the world around the organism which guides its behaviour. It’s that centralised model that produces subjective consciousness. Organisms that respond directly to stimuli in a decentralised way may still produce complex behaviour but they lack consciousness, as do those that centralise the processing but lack the required model.

Traditionally it has often been assumed that the insect nervous system is decentralised; but Barron and Klein say this view is outdated and they present evidence that although the structures are different, the central complex of the insect system integrates external and internal data, forming a model which is used to control behaviour in very much the same kind of process seen in vertebrates. This seems convincing enough to me; interestingly the recruitment of insects means that the nature of the argument changes into something more abstract and functional.

Does it work, though? Why would a model with this kind of functional property give rise to consciousness – and what kind of consciousness are we talking about? The authors make it clear that they are not concerned with reflective consciousness or any variety of higher-order consciousness, where we know that we know and are aware of our awareness. They say what they’re after is basic subjective consciousness and they speak of there being ‘something it is like’, the phrase used by Nagel which has come to define qualia, the subjective items of experience. However, Barron and Klein cannot be describing qualia-style consciousness. To see why, consider two of the thought-experiments defining qualia. Chalmers’s zombie twin is physically exactly like Chalmers, yet lacks qualia. Mary the colour scientist knows all the science about colour vision there could ever be, but she doesn’t know qualia. It follows rather strongly that no anatomical evidence can ever show whether or not any creature has qualia. If possession of a human brain doesn’t clinch the case for the zombie, broadly similar structures in other organisms can hardly do so; if science doesn’t tell Mary about qualia it can’t tell us either.

It seems possible that Barron and Klein are actually hunting a non-qualic kind of subjective consciousness, which would be a perfectly respectable project; but the fact that their consciousness arises out of a model which helps determine behaviour suggests to me that they are really in pursuit of what Ned Block characterised as access consciousness; the sort that actually gets decisions made rather than the sort that gives rise to ineffable feels.

It does make sense that a model might be essential to that; by setting up a model the brain has sort of created a world of its own, which sounds sort of like what consciousness does.
Is it enough though? Suppose we talk about robots for a moment; if we had a machine that created a basic model of the world and used it to govern its progress through the world, would we say it was conscious? I rather doubt it; such robots are not unknown and sometimes they are relatively simple. It might do no more than scan the position of some blocks and calculate a path between them; perhaps we should call that rudimentary consciousness, but it doesn’t seem persuasive.

Briefly, I suspect there is a missing ingredient. It may well be true that a unified model of the world is necessary for consciousness, but I doubt that it’s sufficient. My guess is that one or both of the following is also necessary: first, the right kind of complexity in the processing of the model; second, the right kind of relations between the model and the world – in particular, I’d suggest there has to be intentionality. Barron and Klein might contend that the kind of model they have in mind delivers that, or that another system can do so, but I think there are some important further things to be clarified before I welcome insects into the family of the conscious.