by John Griffin - November 27, 2007
Would you know if you were chatting with a computer program? Could you differentiate [ala the Turing Test] between a computer generated simulated conversation and a conversation with a real person? If you entered into a conversation with an anonymous user who had the ability to parse your language and then spit it right back at you in the form of ‘intelligent’ dialogue, could you distinguish between the robotic response of the machine and the more affective or visceral responses of a real person? Alan Turing, the English computer whiz and decoder of the German Enigma Machine, formulated the eponymous test that sets out the challenge for Artificial Intelligence (AI): He basically posed the question whether a computer program, in contest with a human being, could withstand interrogation by someone who knew it was possible that one of the two interlocutors was machine generated, that is, whether a thinking machine could pass as human. So far the test hasn’t been passed, though a computer contest inaugurated by Dr. Hugh Loebner, called the Loebner Contest, has invited programming participants to submit templates to see if they can successfully create conversation algorithms that dupe the participants into believing they are actually conversing with real people. There have been some interesting results, but before we look at those let’s say a few words about Joseph Weizenbaum.
Weizenbaum was a professor of computer science at MIT. He was a German Jew who fled the Nazis in 1936 and settled with his family in the US. He studied mathematics, and later computers, and developed the program called Eliza, which “demonstrated natural language processing by engaging humans in a conversation resembling that with an empathic psychologist”(Wikipedia). This program prefigured the Loebner Contest, and was the first application of the dialogic patterns we now know as Chat bots (If you wish to know more about Bots, then Andrew Leonard’s book of the same name is the definitive one on the subject). Eliza ‘parodied a Rogerian Therapist by rephrasing many of the patient’s statements as questions and posing them to the patient’. Weizenbaum specifically chose this dialogic situation for parody because he believed:
The therapeutic situation is one of the few real human situations in which a human being can reply to a statement with a question that indicates very little specific knowledge of the topic under discussion.
Not saying much for that kind of therapy, is it? The program worked by parsing and substitution and created the illusion of a human writer in the exchanges. Weizenbaum was aghast at the results of his experiment, particularly at how emotionally involved some people became. From that point on he opposed AI research. Today, one can download an Eliza-Podcast Program and perform one’s own self-analysis. Weizenbaum would affirm that his program failed the Turing Test: the self-modifying stochastic system he invented failed ultimately to convince him of computer intelligence, and he would argue that the ability to simulate intelligence is not itself intelligence.
Many Eliza clones followed, like HeX and SEPO, but the most successful program [MegaHal], devised by Jason Hutchens won the Loebner Contest in 1996. Hutchens' program, the judges felt, succeeded in giving the appearance of conversing with a user in his or her natural language, and it functioned in pretty much the same way as Eliza, by looking for certain patterns in the user’s input and responding with pre-determined output. Transcripts of the conversations generated are available at the Loebner Contest website. The whole point of MegaHal is to construct a model of language that is based on the evidence it gathers while conversing with the user. It parses the input and regurgitates it using what’s called Markov Modelling, a probability distribution system that predicts, selects and applies key words in responses that are aimed to communicate the highest meaning. The algorithm for MegaHal is in four parts and Hutchens lays out what these are in his own online publications. The thing about MegaHal is that it needs to be trained, but it can be trained and, while it generates a lot of gibberish, it is capable of coincidental meaning, as well as causing and giving insult; it can be outed using bad language and can even be used to write ‘poetry’. Hutchens has supplied the text of poems his program generated. Like certain poems I’ve encountered in my own virtual travels, I can attest to the quintessentially doggerel nature of this stuff.
My interest in this topic is itself a kind of circuitous tale told by an idiot. I was researching one Richard Crowley, an Irish journalist, who has recently written a book on the Middle East, titled, No Man’s Land: Dispatches from the Middle East, when my readings took me onto the site of one Dr. Mark Humphrys, a computer professor at the School of Computing in Dublin City University. Once onto Humphrys’ site, I found myself astray in a wild and strange labyrinth indeed. He has many views on many varied topics, and each view comes laden with a myriad hyper-links, so once you enter his lair you invariably find yourself shuttled like a ping-pong from one inflammatory topic to the next. Humphrys is a proud atheist, a devotee of Paine, Orwell, Dawkins, Hitchens, and he holds very far right political views in a country that is still largely Leftist and neutral. He is unequivocally pro-America, anti-Islamist, and thoroughly behind the War on Terror. But it’s not his political notions that captivated me. It’s his belief that he designed a chatbot that passed the Turing Test, or so he claims. Humphrys is a proponent of heuristic programming, or AI, and directs much of his research towards advancing that branch of science. His contribution to the Turing debate was a program he devised as an undergraduate, named MGonz, which generated a conversation with a student from Drake University who thought he was actually conversing with a real person. Jeanne Cavelos wrote about Humphrys’ experiment in the ‘Technology Run Amok’ chapter of her, The Science of the X-Files.
In Humphrys’ version of the creation of his program, he says he installed it in a BITNET chat room and just left it there awaiting its prey. When The Drake U. student logged on he assumed he was chatting with a real person – he never copped on that the chatbot was getting the better of him. Humphrys attributes the success of his algorithm to the element of surprise: this combined with a brutal cross-examination, the repetition of the same question over and over again (as in police interviewing techniques) until the user cracked, bad language, insults, goading and taunts about the user’s sexual prowess all kept him engaged and enthralled even though it is quite evident that the flat, one-dimensional nature of the language suggested a non-human source. Like Weizenbaum and Hutchens, Humphrys realized that people will tend to read more meaning into what is said than is intended, will impose structure on chaotic drivel and cannot help but become emotionally involved. The success of many chat bots today is premised on some or all of these elements. Humphrys talks about the cybersex chatbot created by Jake Kaufman, which was implemented in Eliza. In Kaufman’s own words his program “replaced Eliza’s tiny, boring script with a massive dumb blonde script that has like 3,800 responses on all sorts of topics, but mostly sex. Jenny18 is very horny and she loves talking to horny guys. and everyone knows the best place to talk to horny guys is on dalnet irc sex channels.” The plan is to lure sexual predators into a simulated chat and then frustrate the hell out of them. Humphrys cites this as evidence of a successful AI program.
So all this got me thinking. I know that AOL and Yahoo and Google all employ chat bots using that element of surprise to great effect, and so if it can be used to collect private information, or even elicit deeper intimacies, for marketing purposes, or more nefarious ones as in MGonz and Jenny 18, then why not also use it to lure political subversives? I must confess there have been times on the Fray, especially in ‘Fighting Words’, ‘BOTF’, and ‘Poems Fray’ where I was convinced I was conversing with a such a bot. One in particular stands out above the others, a ‘user’ calling itself AiQuoc. There was no way in my mind this psychotic nutjob could have been anything other than the Jenny 18 of political discourse. But I’ve felt the same about other anonymities I’ve encountered in other fora, and I’ve always suspected that serial plagiarism could be explained by looking at the Markov Modeling protocol of the MegaHal program. And there are other elements to this too, like the ones Weizenbaum repudiated, say, the tendency for people to get too emotionally involved. But how much more entrapped users become when they are verbally abused for no apparent reason?
Imagine you have just joined a discussion board [not unlike The Arena], where you proceed to establish your anonymity with a ‘seen-around-town’ moniker, you get the lay of the land, you weigh and you gauge, you estimate and calculate, and you decide who you need to converse with to gain admittance to the already established community. So, here you are now, after your initial hesitation [and novelty] has passed and worn off, when other anonymities are a little more comfortable with yours, and they begin to engage you and drop their guard and let you into their confidences. And everything is going well with you [and them] until all of a sudden you find yourself under incomprehensible and irrational siege. Out of nowhere barrages of bilious insults and vitriol swamp everything you write. And they are unrelenting, and viciously personal. They come at your from every quarter, and you know all the other denizens are riveted to the flare-up, because people love to slow down near the fatal highway collision. So, you initially try to defend yourself but soon realize, because the vehemence of the attacks is so fierce, that there’s no reasoning with this person, if indeed it is a person. You ask yourself how it’s possible that so much venom could be directed your way by somebody who doesn’t even know you. You say to yourself, ‘I’ve never elicited this kind of savagery in my real life, so what can be driving this person unknown to me to hate me? Or why are they attributing the most insidious motives to my arrival in their midst?’ There is no satisfactory answer to this or any other questions, and none is forthcoming. And yet the character assassinations continue. They even get worse, more irrational, more sinister. And the bot stalks you everywhere you go. Why?
I have a theory about Internet chat rooms as the Hotel California: In many instances users find they cannot break the habit. They cannot avoid participating in a reality which is inimical to their self-worth and their dignity. In fact, there’s a curious co-dependency binding them to their nemeses in these places. Call it one-upmanship, insecurity, an inferiority complex, psychosis, neurosis, or simply wanting to get the last word in, but for some reason they cannot force themselves to leave once the jig is up. But could it be that such places deliberately exploit this weakness and flaw in our unsuspecting natures? What if bots were embedded simply to populate these forums with chatterboxes who cannot leave, convincing them that ‘belonging’ is something they need and crave? How much valuable time has already been wasted consuming the fatuous ramblings of anonymous psychotics and/or computer programs? And can users tell the difference between the simulated conversations of this ‘thinking’ programmed machine and a real person? And if they cannot does that mean then that Turing’s Test has been passed? Does the inability to differentiate imply a cognitive or even pathetic dissonance that will never quite grasp the true nature of the deterministic universe they find themselves in? Maybe, following Wittgenstein’s lead [“An inner process stands in need of an outward criteria”], we might propose a conciliatory reverse Turing Test, and ask ourselves what evidence would we need to verify that this strange and bitter barking person is not just a ‘thinking machine’ with a limited set of incendiary responses?
John Griffin is a writer living in Ireland