Heather McCallum
Professor Ugoretz
Science Fiction: Visions and Universes
1 April 2015
Turing Test
The Turing Test is one of those universalized indicators of artificial intelligence, a two-word summary for a be-all-and-end-all of determinants, just as the ‘Three Laws of Robotics’, a term removing Asimov’s credit entirely, stands in as the summarization of all robotic governance protocols. And like the ‘Three Laws of Robotics’, the Turing Test is ambiguous enough that it may not adequately cover the fullest spectrum of what constitutes ‘intelligence’ and ‘humanity’.
In brief, the Turing Test is designed to determine whether or not a computer could adequately imitate human interaction (since then taken to replicate human interaction, ie, fool the interrogator into believing the computer is human, an important distinction we’ll get into in just a moment) through a brief conversation. Turing, of the era where computers were the size of a room and ‘communicated’ by long strings of ticker-tape, proposed that handwritten messages be passed through from both sides, with an assistant acting as the computer’s conveyance. If the computer was adequately able to imitate human interaction, then he believed it had significant intellectual capacity to be given distinction outside the realm of ‘computers’; it could be granted the distinction of ‘thinking’. As per the wording of the original test, any ‘AI’ that can manage to hold a conversation on any topic without the requisite ‘does not compute’ would manage to meet his requirements. With the right questions, Siri and Cleverbot would make the cut.
What we require of the Turing Test now, however, is replication of human behaviour. The ‘Standard Turing Test’, set apart from the ‘Original Turing Test’, requires that the interrogator be unable to determine whether the computer is indeed a computer or a human. A series of tricky questions and social interactions through a web interface run double-blind with human participants on the other side of the chatbox as often as machines, the end result to be exacted should determine not only whether the ‘AI’ at the test is intelligent, but whether it passes into the realm of what is considered ‘human’. The Original Test doesn’t require that; it makes a point of making sure that all the players are aware that they’ll be taking on personae for the duration of the game, pretending to be something that they aren’t. In the Standard Turing Test, ‘pretend’ is notably absent. The computer is intended only to present itself as human.
In this brief summarization, several points of contention occur. First, the programmer is intended to formulate a program that will have access to a large database of information, potentially culling its knowledge from the web at rates faster than a human could even begin to think to google the request, follow a form of mannerism protocol, and yet in every aspect the entirety of its design is dedicated to fooling the interrogators, to be perceived as ‘human’. Those parameters: several conversations from a series of interrogators and judges, cannot reflect the program’s lifespan. If it is only intended to engage in conversation, and conversation occasionally designed to ferret out whether or not it is a robot, then it doesn’t really have a useful function. Without internet access, the program would have to contain such a tremendous database of knowledge at ever-expanding size and rates that it would be incapable of moving to use as a mobile application, like Siri, or a useful service for businesses. Even accounting for cloud access, as with Cleverbot, such a program would best be purposed toward archival aid, like the library reference desk, or innocuous conversation. (Cleverbot itself builds its interactions by talking with thousands to millions of humans on the internet at the same time, and studying their grammar and responses to repurpose in later conversations. When you have thousands to millions of 12 and 13 year old boys talking about genitalia and pornography, the experiment of a ‘self-building Artificial Intelligence’ fails.) The ‘AI’ cannot ‘feel’ for itself, after all, and genuine emotions are not required to deem it ‘human’ by the Standard Turing Test; only the replication or imitation of emotions is needed. While we accept that sociopaths are human by merit of having a human body and human intellect and being produced by a human through that normal human fashion, they lack that fundamental, integral portion of humanity: not only having emotions and desires of their own, but being cognizant of other’s emotions and desires and navigating those with respect for them. The public is willing to accept that machines have intelligence on par or superior to their own, certainly, but to proclaim a computer ‘human’ with the same emotional capacity as a human? They would revolt.
Alan Turing envisioned an era where the computer would have to pass tests on its visual and physical sensors and servos, essentially looking not only to test computers, but robots. Do Androids Dream of Electric Sheep took this search for ‘humanity’ further to replicating the human body down to the internal organs, and it contains quite a few cautions against trying too hard to crack the Turing Test. To adequately meet the measure of humans, he expected that these computers would have to purposefully introduce error into their interactions, either grammatically, or computationally. He also wrote that the computer couldn’t be smarter than a human, if it were to come across effectively. Both of these things would do well to create a mechanical human, if that were the intent, but would be dangerous to introduce to the wider computational or robotic world. Caps on computational abilities or purposeful and randomized errors would be a serious threat to human life, and could easily poke through the variety of loopholes that the Three Laws of Robotics left. Such a machine could not be globally accepted without wider parameters in place to avoid destruction, injury, or death. Since Do Androids Dream of Electric Sheep came up in the course of this discussion, it is also necessary to point out that instructing the android that it is human, that it must pretend to be human, and providing all the programming basal ‘cognition’ of its humanity is equally dangerous.
Lastly, the quest to fool interrogators provides an end result of a very limited ‘AI’. If requiring a tremendous database but simultaneously capped by an intelligence barrier, the programmer has no choice but to again create a persona for the computer to convey the gaps in its database or ‘knowledge’ without providing the impetus to label it a computer. The single program that has managed to ‘pass’ the Turing Test did so through the persona of a mentally disturbed, foreign 13-year old boy, ‘Eugene’, and it was far from the required or envisioned parameters. It may be easy to blur the distinction where sociopathy, youth, or mental disorders are involved, as it doesn’t require the programmer to instill a healthy complexity of emotional characteristics within the program, but it’s an artificial hobble and a limitation as well. The AI market is not particularly open to mentally disturbed, foreign, tween personae, either. It is clear the Turing Test and its participants have a long way to go.
The future of robotics and artificial intelligence is not going to be based along the introduction of errors or artificial hobbles to present a facsimile of human interaction, but the introduction instead of a diverse range of complex emotions, not only the recognition of them, but the deployment of their own in response to certain external stimuli or stressors, however irrational those emotions may be. And just like humans, these AIs will have to balance emotions and intellect, judging how to respond and how to process two entirely different realms of stimuli. Through the development of these complex responses and a better understanding of human emotions, programmers will achieve a closer match to the Turing Test’s requirements, and help to breach the barrier between intellect and ‘AI’, think and compute, process and feel, react and respond. With the development of the severely limited ‘Eugene’, the Turing Test will need to be carefully revised and reworded to avoid such trickery and interpretation of the rules to receive the best possible results for the upcoming decades.
