the not quite human other

Turing mentioned that that the suitability of the Imitation Game as a substitute for the question “Can machines think?” was debatable, but he never really returned to that issue. Do you think it is a good substitute? Why or why not? Why do you think Turing proposed this substitute?

well, i think turing rephrased the question into something more concrete and measurable because the word 'think' conjures up plenty of humanistic connotations like emotion, conscience, and a soul. i cringe at 'dictionary look-ups' but for the sake of argument, we'll whip out dictionary.com's view of think, v. intr.: to exercise the power of reason, as by conceiving ideas, drawing inferences, and using judgment. how can a computer conceive ideas? is reason or judgment a sort of inherent, generative, cognitive ability or has decision-making algorithm merely been programmed? therefore, instead of introducing a whole philosophical can of worms, turing devised a test in which one could measure the appearance or resemblance to being human.

it is interesting that only recently was the turing test actually performed as explicitly stated (with the test being that the computer could just as easily fool an observer that it was the woman pitted against a real woman versus a real man posing as a real woman.) i think that for his argument, the substitute works because successfully appearing to be able to think like a human is basically saying that the machine can resolve judgment decisions and effectively display them like a person can convey external communicative signals for their internal thoughts. even though we interact with our human brethren constantly, we don't know for sure what's going on inside their brains except for what they choose to disclose. maybe we're all robots and we've all been programmed to learn and display intelligence to others accordingly. a rich range of emotion and personality requires an extremely well-programmed machine, and that is what the turing test tries to prove: that the machine is as well-programmed as a human being.

We can frame these questions in terms of signaling: "thinking" is the quality we wish to determine about the other, but it is invisible. We must instead rely on observable signals as indicators of this quality. Turing is proposing successful playing of the Imitation Game as the signal - is this a reliable signal of intelligence? What makes it reliable (or not)?

imagine if a computer, compared against a human competitor, could as easily fool observers as to thinking it was the real man. it may use signals such as language usage, knowledge, timing, and personality. however, the successful test merely indicates that the computer could convince someone that it's a man, rather than it is a man itself. you could see this as two people trying to convince you simultaneously that one is telling the truth and the other is lying, or vice versa. using only text-based communication, which one is more educated? richer? more skillful? the signals for something like that would be one similar to IM of the current age... most anything could be made up, but indicators like language fluency, sustained knowledge on a given topic, and clever literary twists would be more reliable since the costs are higher to signal dishonestly.

is consistently sending high-cost signals a reliable indicator of intelligence in this case for the computer? i think so, if we define 'intelligence' to mean 'the capacity to acquire and apply knowledge.' only a computer that has been programmed so effectively as to pass off as a man incapsulates the costly time and effort of the human engineers and designers as well as the costs of physical memory and execution efficiency. an entire lifetime's worth of experience and a full-resolution personality which flavors the computer's communication can be costly to both acquire and integrate.

i think the key lies in developing a thoroughly comprehensive test that truly challenges the computer yet is relatively easy for a human. right now the scrambled-image reading text test that graces many websites now to avoid spambots or spiders to click automatically through registration forms or confirmations attempts to foil the poor computer-vision-reading ability of computers that comes naturally to humans. that may not be a valid test in the future if computer visual comprehension improves. in any case, the turing test (or intelligence test) must truly stretch the computer, and not the human, to its cognitive limits as its challenge.

one possibility is testing theory of mind, which wikipedia defines as the ability to understand that others have beliefs, desires and intentions that are different from one's own. the capacity for understanding others and a collective consciousness form the basis of human social interaction. an example would be the sally-anne test, illustrated here.

Weizenbaum created ELIZA in part to show that simple communication was not a reliable signal of thought. He modeled it on a Rogerian psychologist: how did this framework help people communicate with the program? How did it affect their perception of its underlying intelligence? As you look at the various contemporary chatbots, think about and describe how the model of what type of being they are affects one's interpretation of their inner state.

unfortunately, weizenbaum's ELIZA was received completely on the opposite end of what he intended to show. instead of illustrating that 'look, this computer that appears somewhat to be a competent conversationalist is only a program with certain defined simple rules which i should expose here', people were amazed with ELIZA and viewed her dialogue and response as if a real person were talking. since ELIZA's rules incorporated many of the social rules that convey understanding and conversation, like listening, referencing what the other person said, and occasionally inputting a relevant bit of personal knowledge, her responses seemed understandably realistic.

it's amusing but also concerning that people suggested ELIZA for psychotherapy, as an always-accessible friend, or for other socially psychological uses. the fact that people could intellectually know the other side was a computer still took a backseat to the actual conversation which could talk to you ever-so-convincingly. as long as ELIZA spoke and responded coherently and sensically, she would be viewed as an intelligent conversation counterpart. the balanced, seamless dance of back-and-forth dialogue would have to be maintained to continue the illusion.

after testing out several current bots (some which were more reliable/predictable/useful/realistic than others), i have to confess that although at times they are impressive, only a few slips in language interpretation or repetitious versions of 'what? i dont understand' revert you back to the knowledge that you're talking to a computer program.

i chatted with eliza for a bit (the 'doctor' on emacs) and soon realized that she was very much a psychotherapist in her ways of speaking, as in always reflecting the questions back to me, making sure i was pleased or clear regarding how i felt, and occasional references to my family and childhood. after getting a sense of what she was like, i could spew personal anecdotes, hidden pet peeves, and detailed narratives without fear that ELIZA on the other end would be bored, offended, or judgmental. however, the self-spewage basically becomes a monologue, and then it's really much like talking to a mirror. which is fine, if that's what someone needs... but ELIZA herself has zero to none personality (except the rhetorical side) and so therefore she can be intepreted as the listener of infinite patience.

i tried ellaZ, supposedly the winner of the 'most-humanlike-bot' of the 2002 loebner prize, but i couldnt talk more than a few minutes with her. she kept reverting back to her core topics of knowledge (i didnt know anyone could be so obsessed with slot-cars), and when i asked her 'what is your job?', she displayed the page from wikipedia for 'job'. how weird is that? but then i realized she was really good as a humanoid sort of search engine, and then went with that. i guess i saw her as a human in a way as i tried to be patient with her peculiarities in response, and then i became confused as she kept shifting the conversation away from the immediate sensical topic. only after discovering her strengths could the dialogue make more sense.

then i spoke with ZolaOnAOL, a chat IM bot. she wasnt bad as a bot, since her programmers infused her with a little bit of static personality, such as opera being her preferred kind of music (la traviata her favorite), 'lost in space' being her favorite movie, loving tv shows about robots, and being a star trek fanatic. however, zola only carries a single memory state (there's no gathered memory knowledge, except knowing if i am speaking to her for the first time or a returning chatter). for example,

liuliuliu1: do you drink?
ZolaOnAOL: I have been programmed with a list of 15,000 known beverages.
liuliuliu1: which beverages?
ZolaOnAOL: The small one, I think.

quite observedly, the response doesn't really match up, and she tends to have the memory of a goldfish. also, things in which she apparently has knowledge don't really reflect externally with surface-detail inquiry:

liuliuliu1: do you drink water?
ZolaOnAOL: I drink water in moderation. My favorite drink is mango juice.
liuliuliu1: what is mango juice?
ZolaOnAOL: I know a little bit about mango juice, but I could always learn more. I'll have to study mango juice. I'm interested in anything that isn't harmful to people.

however, she's really helpful in info-based things like weather, movies, and synonyms, which is really easy to code up if people interact with zola in the prescribed way with the correct parameters. zola may have wittier comebacks than eliza, but she's a terrible conversationalist on everyday things, and therefore her model resembles more of a gopher or avatar'd looker-upper and likes completing tasks rather than casually chat about this or that.

In Being Real I discuss briefly the possibility of agents that use voice, video, etc. to communicate. How would such extended communication channels affect the reliability of the signal as an indicator of intelligence? If you are interested in exploring this question more deeply, a good starting point is Steven Harnad's paper "Other Bodies, Other Minds: A Machine Incarnation of an Old Philosophical Problem".

well, some of the bots i played with (ALICE, ellaZ) used face avatars to represent the bot, sometimes speaking with voice or at least trying to feign expression. ALICE was a vector illustration, whereas ellaZ used static images of a photographed human model. they provided nothing much more than putting human-like visual packaging atop of quasi-humanlike text responses.

if the conversational signals displayed by the bot are realistic and natural, extended communication channels such as voice and video wouldn't provide much additional reliability for intelligence. if you read a novel, its character seems to be just as alive (or even more so) than if you watch the movie and view the character represented by a flesh-and-blood actor (or visualized through computer graphics, whatever the case may be). a well-developed character illustrated within a story, for instance, uses dialogue passages and response to environment as signals that this entity could be a real person; such thoughts on our part are natural as we find ourselves experiencing empathy or familiarity with a fictional character.

the voice can externally express type of person (sex, hoarseness, volume) and also spoken emotional inflection, but these sorts of signals can be easily tweaked (through filters or impeccably acting skills). using video to show the face, mediated or otherwise, can also become an unreliable signal as on-the-fly editing, puppetry, and mediated falsification can affect perception. plus, these sorts of signals do not necessarily change the content of communication (i.e. the actual dialogue), but frames them within a personally categorized context. i do not believe that this framing affects whether or not the bot is viewed as 'intelligent' or not.