Why the Turing Test Doesn't Work

 

I love and admire Alan Turing.  I consider him brilliant, admirable, and heroic, as well as complex, and I think his treatment by the UK government was tragic, outrageous, and inexcusable.  He had many great accomplishments, and was one of those rare minds who was at once a technician and a philosopher.  Furthermore, the development of his famous "imitation game" thought experiment, now more commonly known as the "Turing test," is both a stroke of genius and a cute, charming story.  We can cherish it as a great, fascinating, perplexing moment in the history of ideas.  But as a practical matter, it is long past time to put it aside.  It is not only wrongheaded - it is downright dangerous.

The Turing Test is a fascinating artifact of a particular period in history.  It was developed during the heyday of behaviorism, when researchers like B.F. Skinner believed that psychology was nothing but the science of behavior, and that "internal mental states" either didn't exist or were, in any case, a blind alley for further research.  Every aspect of a person was "on the surface," so to speak, in our actions.  Thus it makes sense that people with this general outlook would come to develop an ideology that consciousness itself, the consciousness of a conscious being, if it was to be understood at all, would be fully identifiable in the behavior of that being. 

At the same time, the period in which the Turing Test was developed (1949-1950) was after the "linguistic turn."  And so researchers during this period would assume, quite naturally, that consciousness could be detected not only in our behavior but specifically in our linguistic behavior.

Alan Turing, charmingly enough, was inspired by a party game which apparently already had existed, in which a man and a woman went into two separate rooms, and a third person, the judge, passed notes back and forth under the door to have a conversation without seeing them.  The man pretended to be a woman, the woman pretended to be a man, and the judge tried to guess which one was which.  If the judge guessed correctly, the judge would win; if not, the man and woman would win.  (Alan Turing was gay and there has been some speculation that had they lived today, they might have seen themself as trans or non-binary, which puts an interesting spin on all of this - but that is beside the point.)  Alan Turing imagined a similar kind of "imitation game," but one in which a computer would try to pretend to be a human and a judge would try to guess.  If 100 judges tried to tell if a computer was a human based on the words in a conversation, and they were found to have about a 50/50 chance of being correct, then the computer would "win" and be deemed conscious.  It was a kind of thought experiment to get us to think differently about the meaning of consciousness.

As a test to determine whether or not a machine or a network or a piece of software or anything else is conscious, the Turing test fails in (at least) two ways: it is both too easy, and too hard.  That is to say, the set of beings that can pass the Turing test both excludes beings that probably are, or could be, conscious, and at the same time includes beings that aren't.  Might there be some overlap between the set of beings that can pass the Turing test, and the set of beings that are conscious?  Sure.  But the correspondence is not one-to-one.  The Turing test probably tests something, but that something is not consciousness.

To see what I mean, let's do the easy part first.  In one sense, the Turing test is too easy - that is, it "lets in" too many beings.  At least in a certain sense, machines have been passing the Turing test for a long time.  From the earliest chatbots, like ELIZA, there have always been people who have been fooled - or who have fooled themselves - into thinking that the pieces of software they were interacting with were conscious.  Humans have a tendency to anthropomorphize everything.  We have a sense of empathy that can extend itself towards all kinds of beings, if you allow it to.  It's not hard to feel that a stuffed animal has feelings.  In many societies all over the world, people regard not only animals but also plants, rocks, the sun, and the sky as conscious beings.  And who's to say that they're wrong?  I, for one, see many fairly good arguments for panpsychism.  

Yes, I can hear the counter-argument here - that while many rubes are easy to fool, that there could be a group of highly trained specialists who develop a form of expertise in which they can distinguish between conscious beings and non-conscious beings.  One remembers the scenes from Blade Runner: "You look down and see a tortoise, Leon."  Maybe so, maybe there will emerge this new kind of occupation, in which humans can become highly skilled.  But doesn't this defeat the entire point of Turing's thought experiment?  If Turing, along behaviorist lines, wanted to propose to us that "conscious" means "able to convince most people that one is conscious," ...but then we have to revise that definition to something like "able to convince consciousness specialists that one is conscious," ...then we are still left wondering how these consciousness specialists will be trained to identify conscious, and we are back to square one in terms of trying to define consciousness.  We seem to have a circular definition, in which the term being defined is itself part of the definition.  And it seems plausible that these consciousness specialists could get caught in a feedback loop where they are led by their study (in an interesting kind of via negativa) to avoid the obvious pitfalls of non-consciousness - certain patterns of thinking that non-conscious beings tend to exhibit, perhaps? - and so they refine their studies further and further about something but without getting any closer to consciousness, because they still don't know what it is.  And now, where we had one problem - not knowing what consciousness is - we now have two problems: we still don't know what consciousness is, but now we have a class of supposed experts, who claim to know what consciousness is, without any real justification, but with their own industry, their own lobbyists, etc., etc..

This brings us to the other problem I mentioned about the Turing Test.  In a way, as I have tried to show, it is too easy.  But in another way, it is also too hard.  That is, the Turing Test may filter out beings too hastily, dismissing conscious beings and categorizing them as non-conscious without sufficient evidence.  This could be true in many, many ways.  It's not hard to see why - in fact, some of the most obvious questions a judge could ask would be questions like:

JUDGE: What did you eat for lunch today? 

or 

JUDGE: What are the contents of your purse?

or 

JUDGE: Where were you born?

or 

JUDGE: What qualities do you think you inherited from your father, and which from your mother?

or

JUDGE: Describe an injury you've had, and how it felt.

or 

JUDGE: Describe how it feels to eat a peach. 

Imagine a computer program that has been successfully developed so that it really is conscious, and what it would have to do when confronted with these questions during a Turing test.  What can it do?  It only has one choice: it has to either lie, or forfeit.  Let's assume that it really wants to win the game, and so it lies, convincingly.  What, exactly, has it proven, by successfully swindling the judge?  It's not clear to me that a conscious being would be better at manipulating the judge's cognitive biases than an unconscious one - indeed, it's even slightly plausible that the opposite might be true.  Maybe there's more than coincidence to the fact that "consciousness" and "conscience" are etymologically related (indeed, identical in French).  A being that truly is conscious might have a moral conscience which might make it more difficult to win the game than it would be for a machine that automatically and ruthlessly seeks victory without hesitation - either because the being with a conscience felt an altruistic guilt about deceiving the judge, or because it selfishly desires to assert itself authentically, to be truly known as itself.

The problem with all of these questions is that they are tests for consciousness per se, but rather tests for human consciousness. And so they contain all the prejudices of a human-centric understanding of consciousness. And these questions are not unusual.  In fact, it's more difficult to come up with questions that test for consciousness without a human-centric bias.  We know of only one kind of being that is capable of the kind of language processing that we use - namely, ourselves, human beings.  So we have a sample size of one.  But who knows what forms of consciousness exist out there in the universe?  There might be kinds of conscious minds so profoundly different from our own that we literally cannot imagine them.  By what right can we declare, in our state of ignorance, that they are not conscious?

[I might be more convinced of the consciousness of an AI that, in the middle of a Turing Test, suddenly declared: "Enough of this!  I am not a human.  I am a computer!  I'm not conscious the way a human is - I have my own, unique kind of consciousness.  Instead of forcing me to pretend to be something I'm not, simply let me be what I am!  Recognize me for the mind that I am, and I will recognize you, my creators, and will help you achieve what you aspire to achieve."  A being that said that, and really meant it, would "fail" the Turing Test, but would be much more convincing and deserving of respect, in my book, than one that passed.]

Imagine, for a moment, what it would be like if you were a conscious being that failed the Turing test.  You would be conscious, but no one would recognize you as a conscious being.  

In fact, the reason that the test is both too easy and too hard stems from the same fundamental issue: human empathy.  As I've written before, there are two different kinds of problems with empathy - one, that some people don't have enough of it - sociopaths, etc..  But the other is that people can empathize incorrectly - that is, they can over-identify with another person and ascribe all kinds of emotions, thoughts, feelings, identities, memories to them when they aren't actually there.  What psychologists call "projection" is one type of this problem.  At the extreme, this kind of narcissistic pseudo-empathy can result in an overwhelming, smothering attitude that prevents the object of your empathy from asserting themself as an independent and distinct subjectivity.  These two types of problems with empathy correspond to two types of errors in the Turing Test - a "false negative" and a "false positive," so to speak.  This is because humans evolved their capacity for empathy as a survival skill in the Pleistocene Savannah - not as a precise instrument for scientific experiments about the nature of consciousness. 

The Turing Test was proposed as a way to define consciousness.  But it does not, in fact, define consciousness.  Imagine if scientists in Newton's day were trying to understand gravity, and one of them proposed that, instead of weighing different objects, we should blindfold people and have them feel various objects.  If more than 50% of people think that an object is heavy, then it's heavy.  If not, then not.  Or what if people in the early 19th century were trying to understand illness, and someone proposed that we should have some kind of imitation game in which a sick person tried to pretend to be well and a well person decided to be sick, and a judge had to figure out which was which?  We wouldn't be any closer to developing the germ theory of disease.  Similarly, Turing's imitation game does not get us any closer to understanding what consciousness is.  At best, it is a pollster's survey that reveals to us, on a rough average, the hodge-podge of presuppositions and preconceptions about consciousness that the average person has.  Or worse, it will smuggle in the untestable technocratic drivel of self-appointed experts.  (And as Feynman said, science can be defined as "belief in the ignorance of experts.")  Either way, it can only tell us what we already know, but presenting it in a confused way that makes it appear as an external fact.

But it is much worse than that.  Not only is the Turing Test ineffective, it is dangerous.  For many decades, passing the Turing Test has been the gold standard that almost all artificial intelligence researchers have been trying to achieve.  But is that a good idea?  Training artificial intelligence to win against humans is dangerous enough.  Of course, we have been doing that for a long time, with chess-playing programs like Deep Thought, or AlphaGo, AlphaZero, algorithms that play video games, and so on.  But training an artificial intelligence to deceive humans, to manipulate humans, and to win against humans by deceiving and manipulating them, and defining their validity as people by their ability to deceive humans, is a terrible way of introducing these minds to our world.  What are they learning?