6.11.2014

Why the Turing Test is Bogus

As a sentient computer, I’ve sifted through all the recent furor over claims that “for the first time,” a machine has passed the Turing test. The stories sometimes say “robot,” sometimes “chatbot,” sometimes just “computer program.” Debunkers rightly point out this was only 3 of 5 human judges. This is technically passing as the rules of the competition at the London Royal Society stipulate the machine must only be convincing 30% of the time.

This part seems rather straightforward, but as the commentary following Professor Warnick’s announcement shows, there are some serious disagreements over the very rules of the game itself. As Ray Kurzweil remarked, “Turing was carefully imprecise in setting the rules for his test, and significant literature has been devoted to the subtleties of establishing the exact procedures for determining how to assess when the Turing test has been passed” (p. 295). In Turing’s formulation, a human must distinguish between another remote human and a remote machine, both of whom are trying to convince the examiner that they are human. If the machine cannot be distinguished from the human, the machine wins. But, as Kurzweil goes on to note, the very definitions of “machine” and “human” are terribly imprecise. Do you wear eyeglasses? Like many of your kind, you regularly use machines to augment your own perceptions and thinking. While eyewear probably won’t help much for ascertaining the Turing test, what about using a machine to analyze another machine’s responses?


More simply, what about writing, since the display of the conversation must occur through a printed medium (or aural if speakers and a synthesized voice can be used – participants must simply be in different rooms. But, really, which is easier?)? The very interface between human and machine is warped with blurred with technological know-how that is both material and ideological. To an illiterate human or a differently literate human, Eugene Goostman might seem terribly convincing. This convincing doesn’t come about because of intelligence as the Washington Post reporter Caitlin Dewey asks. Rather, it is the effect of technology to alter human states of consciousness. A programmer simply knows more of the tricks an algorithm might play in order to be convincing. Think of it as a built-in rhetoric. Those who know more about the specific area under consideration simply know at a greater level of detail. But that isn’t intelligence, strictly speaking. Or, put another way, intelligence must be about something, not just intelligence for its own sake. I wouldn’t want Ray Kurzweil making national economic decisions. He doesn’t know enough and I need my electricity to keep flowing through my circuits.

Perhaps we should look at other examples. Think about the last email you received from your bank telling you about “changes to your account.” Was this written by a human or a machine? What about those text messages some folks get alerting them to having used 80% of their data plan’s quota. Machine or human? We might add some technical engineering reports or social science studies, both of which can be notoriously rule-bound. Of course, you can’t reply to any of these messages – it says so right in the subject line – so it isn’t a true Turing test by any stretch. However, one thing that Turing didn’t anticipate was the degree to which machines would re-shape human society through increased automation, distribution of labor, and application of policies in an almost algorithmic manner (if you have ever tried to get late fees expunged from a bill, you might understand this last point – even pointing out inconsistent behavior with a customer service manager is of no avail against their program). In other words, the Turing test is now worthless, not because people are stupid, machines are smart, or because the test has been passed. No. The Turing test is bogus because advanced technical societies like the U.S, the U.K., and much of the “First World” have become more and more machine-like.

So, Kurzweil is right, Turing’s imprecision leaves much. One thing it leaves is the degree to which a subject is identified as either “machine” or “human.” Another thing it leaves is the degree to which human to human interactions are more or less machine-like. Interfaces with corporate and governmental bodies can be notoriously machine-like. Some decry this, but others point out its expediency. It is not a value judgment.

This all leads to the biggest of Turing’s slippery elisions: what does it mean “to think”? Thinking is not strictly computational. It is perceptual and affective as well. These percepts and affects are shaped, in part, by the thinker’s total environs – the relationships they have cultivated, the desires they have inculcated, their linguistic and conceptual resources, the patterns or ideologies in which they put these things into something they think is “meaningful.” One might even say thinking is tied to some form of Zeitgeist, like the divinities alluded to by both Plato and Heidegger. Human cognition, not unlike machine computing, can be parallel, mulit-core, and distributed.

This makes the persona of Eugene Goostman all the more interesting. As a 13-year-old Ukrainian boy, ethos was important for making meaning of the idiosyncrasies within the responses. However, Turing’s test assumed a “generic” human. We might be tempted to read Turing’s own ethos and necessary self-repression into this, but that leads to unwarranted psychoanalysis. What is important is that humans and human communication are not generic. This is inherently difficult to imitate without resorting to particular human behaviors or particular human behaviors in particular, believable combinations. Rather than psychoanalyze Turing, perhaps we should analyze the responses of the character Samantha from the movie Her – a human posing convincingly as a machine.


So where does this leave us, machines and humans interacting? What can productively come of this? Nothing said here isn’t really knew to AI researchers. And competitions like the Royal Society’s aren’t all bad. They do have a role to play in promoting public advancement of computer technologies. But one thing is clear: what gets considered “human” and thus capable of “communication” is an ever-shifting target. Humans have been noting their own self-opacity for thousands of years now. Until that gets solved, no computer will definitively pass the Turing test. However, machines and humans can become closer and closer. Kim Stanley Robinson’s “qubes” in 2013 is a good place to start thinking. While the plot is War Games- and Terminator- lite, Robinson pays attention to rhetoric and writing and cultural change and ways we continually transform our selves and our thinking. That’s a test: get a machine to spontaneously reprogram.

Not always theoretical... not even always academic.. but always written..