by Jeremy Douglass | jdouglas@umail.ucsb.edu
ENGL236: Hyperliterature | http://transcriptions.english.ucsb.edu/courses/liu/english236HL/index.html
Professor Alan Liu | ayliu@humanitas.ucsb.edu
University of California, Santa Barbara English Department | http://english.ucsb.edu/
Fall 2001
Against Machine Writing: A Historical Sketch
Against Queer Identity: A Personal Perspective
A Biographical Sketch of Alan Turing
Against Turing Tests: (mis)Applying the Theory
Passing through Passing: Intelligence Process vs. Intelligence Effect
From Intelligence to Machine Writing
From AI Impersonation to Agents: How the Turing Test fathered Bots
Towards Writing the Future: (un)Conclusions
Machine writings are texts produced mechanically in accordance with some composition strategy. [1] These texts need not be in a traditional literary form, although for the purposes of this discussion they will be assumed to be in a common style familiar to a human reader, such as a monologue, poem, letter, short story etc. The machine itself need not be a computer [2] – although as with many large and complicated problems of the past century, the computer is currently the only machine capable of manipulating the rules and components of language in anything approaching the complexity used by human writer.
Mathematician Alan Turing (1912-1954) outlined the Turing Test, or “imitation game,” in an article [3] published four years before his death. The test could take many forms, but essentially consisted of an interrogator at a text terminal. The terminal was used to communicate with two candidates, one a computer attempting to pass the test and one a human control. [4] The computer was considered intelligent if it was able to fool a judge of the conversation [5] into considering it human for a significant length of time or a significant percent of the time. [6]
By refusing to create a list of specifications for how the computer program would function or even define precise criteria what it should be able to accomplish, Turing created a compelling and lasting benchmark for the interaction of a computer text system with a human reader [7] – a benchmark which I assert that race (and later queer) theorists already termed ‘passing’ when describing human-human interactions. For a computer, ‘passing’ as human is all that is required for passing the Turing Test.
This paper considers machine writing and its future in the context of Alan Turing’s test and the impact of that test on the field, both philosophical and technical. I consider the Turing Test broadly in terms of articulating a theory of art, identity, and human experience. Read in this way, machine writings become ‘authored’ by the machine in the moment that they are no longer easily distinguishable by humans from other human writings – when they ‘pass’ as human. This conclusion is interesting because:
Before sketching the Turing Test in more detail, however, I want to set up the debate on machine writing and authorship to which it would contribute – particularly in terms of sketching the ongoing humanist reaction against technologies.
The late 20th century Information Revolution brought us a narrative of the human vs. computer that was similar to the earlier Industrial Revolution narrative of the human vs. factory the power loom or the mill. These two historical transitions are the latest steps in one narrative of human technology, in which the slow pressure of increasingly powerful human tools pushed back against the human skills which those tools embodied, systematized, and in some sense perfected - a narrative of antagonism between humans and their externalizations/creations.
During the Industrial Revolution, humans retreated from claiming superiority in the manual production of goods to claiming superiority in “skilled” labors (which were often negatively defined as those labors which humans were better at then machines). These included among other things the use and production of languages and the fine arts. Human technology had circumscribed and limited the domain of human superiority at the same time it had extended the power of human civilization.
The Information Revolution brought a similar conflict with the computer, which again it extended human power while limiting even more sharply the areas of human superiority over human-wielded tools. Increasingly almost any activity that could be reduced to an operation could be done digitally with far greater speed and accuracy. [12]
Machine writing is the latest and perhaps one of the most cutting assaults into the heart of what has hitherto remained human domain – the arts, particularly writing. Certain inroads have already been made, to be sure. A computer can take a photograph, selecting the framing of its composition and balancing its color and contrast. A computer can model, specify, and produce an abstract sculpture. A computer can even ‘look’ at a hillside and render it its view into a pointillist painting or a watercolor. However, while some techniques of production have been surprisingly accessible, complex symbolic manipulations (such as the composition of music, or of essays) have remained largely closed to the computer thus far.
Yet human eminent domain continually falls back, retreating in towards the analogical and an increasingly narrow ‘artistic.’ Spelling is no longer the provenance of humans, but grammatically correct language remains – in part. When that fails, topically coherent language remains… then artfully rendered and socially appropriate language… then high language… then great literature?
This is a particular political narrative of antagonism between humans and their technologies, and I’m considering machine writing in this context for several reasons. Machine writing is the latest and perhaps greatest transgression against all things definitively human, and thus its historical development and current aesthetic categorization cannot be treated as abstract philosophical exercises. This transgression however stands in direct opposition to the progress narrative, in which the ostensible project of all technologies-as-tools is as ‘natural’ extensions of humanity and of human works. While the transgression and progress narratives stand in direct opposition to one another, the progress narrative applied to most technologies is not particularly compelling when applied to machine writing [13] – machine writing does not solve a general human problem, meet a great societal need, or satisfy a particular market demand. In this, unlike the power loom or the calculator, a machine that writes sonnets or short stories is an externalization of part of what seems particularly human for its own sake. It is ‘pure’ humanities – and thus often seen as pure transgression without mitigating socio-economic remuneration. [14]
Such transgressions however were the stuff of Turing’s work – a condition not only of his philosophical bent but of his life and eventually his death. [15]
Much has been written on the life of Alan Turing. Most of the material focuses on his role as a mathematician, an early computer scientist, and an artificial intelligence pioneer. Some looks at his work in cryptography, and his role as a decorated British hero in the cipher races of WWII. Some considers his identity as a gay man, a convicted sex offender under England’s oppressive post-war anti-gay ordinances, and once convicted the subject of experimental hormone treatments that led up to his suicide. The fact that Turing was a brilliant and somewhat eccentric mathematician is interesting background information, and his role as an influential pioneer in such emergent and bifurcating fields as sequential analysis, computer science, cybernetics, neural networks, and artificial intelligence helps us to understand the sheer social impact of his theories on the present-day culture of doing machine writing. However many biographers have elaborated this history and authority better than I could, and I will only refer and defer to their research and conclusions here. For the purposes of this paper Turing’s position of scientific influence will be assumed, and we will focus on Turing in terms of his understandings of identity, examining the slippage between personal and theoretical positions, using queer identity as a way in.
My primary source for discussing Turing’s life and particularly his queer identity is Andres Hodges excellent biography Alan Turing: The Enigma. [16] I re-present Hodges work here, although I differ with Hodges conclusions these experiences generated confusion which was expressed as a gender “red herring” in the Turing Test. [17]
Turing is not untheorized as a queer figure. His alienation yet his insistence on remaining out with his friends, his betrayal by a former lover, his detailed confession of homosexuality but not of wrongdoing to the police, [18] and his defiance in court make him a striking figure of quiet resistance during a wave of anti-gay legislation that swept England. [19] His conviction was treatment with estrogen injections, [20] which it was hoped would diminish and eventually destroy the sex drive, thus removing his ‘criminal’ desires. Turing himself accepted the treatment rather than the jail time because he refused to believe that desire was purely biochemical, or that identity could be manipulated by such crude methods. Writing to a friend, he speculated that his physical desires would return once the treatments stopped – his emotional ones had not changed.
That gender and sexuality, love and desire were always intertwined with Turing’s work is an argument that an argument that has been made in many different ways. A favorite example is Turing’s work with Christopher Strachey in 1952, in which they used a random number generator on a computer to compose ‘love letters.’ One output:
Darling Sweetheart,
You are my avid fellow feeling. My affection curiously clings to your passionate wish. My liking yearns to your heart. You are my wistful sympathy: my tender liking.
Yours beautifully, MUC
(Hodges, 477)
This has been discussed in terms of legitimacy of desire, and fascination with / alienation from love, however for this paper I must return to reception – the true message of this love letter “this is a love letter.” The words selected and assembled, the synonyms recombined are all merely strategies of representation – an attempt to evoke what the program termed “avid fellow feeling.”
Another common example in Turing’s work is gender-as-cipher – the odd terminology adopted by Turing’s war-time office to refer to the fundamental decryptable unit of German code as a “female.” [21] Much has been made of this, with explanations forwarded having to do with the punching of holes in cards, matching versus unmatching X and Y chromosomes, and the inevitable female-as-enigma – the German encryption device was, incidentally, known as “The Enigma.”
However for a man who wrote theoretical essays speculating on what it meant to be human, or to act human, the narrowness with which others viewed the human condition and its “natural” states was the primary negative determinant of his life. Although he both new the game too well and refused to play, life itself was an imitation game for Turing, [22] in which even as a human he was forced to “pass” as such, convincing interrogators daily of his human status in accordance with their preconceived criteria rather than his own. The “imitation game” of his Test both subscribed to and parodied ideas of gender difference by positing a man and a woman trying to “pass” as female to an interrogator. Critics later lambasted the sloppy analogy and its failure to make clear the issues of intelligence at stake in such a test. After all, what if a man was better at “passing” as a woman than a woman was? Turing’s understanding and experience of this however was all too visceral – as his sentence to estrogen injections continued, his body developed breasts. [23]
The day he committed suicide, he soaked an apple in a concentrated cyanide solution and then took a large bite. He was found lying in bed with the apple lying next to him on the nightstand – a fairytale image.
His mother never accepted the suicide, insisting that the apple had been exposed to cyanide accidentally. After ritually warning Turing to wash his hands of laboratory chemicals for years, she seemed to finally be witness to the accident she had predicted and feared. For Turing, it was the perfect crime – the suicide he desired while leaving a symbolic void upon which his mother could project her own interpretations in accordance with her own expectations. When she wrote his biography, she reported his death by misadventure. [24] [25] [26]
The Turing Test was an interrogator interacting in a limited with candidates attempting to “pass” the test by “passing” as intelligent for some significant length of time. Turing posited this intelligence not as an intrinsic quality, but as something granted once the interrogator rendered judgment in accordance with an arbitrary set of criteria. I’ve tried to frame this as an assertion made by a man who had experienced how appearances (and the reception of appearances) functioned in the construction of gender and sexual identity. While this does not prove that he harbored some hidden program of postmodern identity that I must now elaborate for him, I hope it indicates the position from which he theorized, and perhaps the sympathies of his theoretical productions to my current program of appropriation.
The history of applying the Turing Test to actual projects is also one of contention and resistance. From the beginning, Turing anticipated many objections to the idea the machines could think, which he addressed briefly including.
Turing raises these objections in order to debunk, and he manages fairly well – some (such as points 2 and 4) are almost comic in the disdain he holds for them, while others (such as 3, 5 and 7) present some serious difficulties. I want to acknowledge here that the Turing Test is intimately tied with this discussion of the possibility of machine thought – machine creativity, for my purposes – yet admit that I cannot give these debates fair treatment here. [37]
Fierce debates in the fields of artificial intelligence and cognitive science since Turing have led many to call for a complete rejection of the Test as a standard, arguing that this act of ‘passing’ by a program does not in itself measure intelligence at all. In 1981 Ned Block argued that a system could pass the test without acting intelligently – his argument ironically hinged on the definition of intelligent behavior, which Turing had attempted to circumvent through developing the test itself.
Block chose to interpret Turing’s assertion (that intelligence was the ability to appear so) as synonymous with “only an intelligent system can give responses that appears intelligent.” However Turing knew this not to be the case. Rudimentary systems can appear to display a high level of intelligence given a limited context – a side effect of personification and the general mystique of new technologies which has always haunted computer scientists, and which has been dubbed “The Eliza Effect.”
Although Daniel Dennet in 1985 and John Moore in 1987 argued that the test was a “quick probe” of “deep” intelligence, and neither too brief nor too narrow (pointing to the vast scope of possible testing activities), they fell into the same basic trap of trying to claim that the expansiveness of the test reflects on the machine rather than the reader. Yet, according to Turing, his test was an “imitation game,” and intelligence was the ability to appear so.
The other half of this argument against the test was that an intelligent being could fail. Examples were a space alien, intelligent animals such as a gorilla or dolphin, a small child, an uneducated person, and even a bored or asocial human. Even a poor typist or slow conversationalist might fail in competition with a Chatterbot in an English-administered test lasting only a few minutes. While unlikely, it is certainly possible. Even more damning is the cultural critique – even an English-speaker reader of this essay might fail the Turing Test miserably if it was administered in Chinese. A highly educated Chinese speaker might be unable to type in Chinese – possessing intelligence but not a method of impressing that intelligence on the Turing interrogator. If these intelligent beings could fail, couldn’t an intelligent machine fail the Turing’s Test, thus rendering it useless for proving intelligence?
Searle [38] and other cultural critics of the test interpreted Turing’s assertion (that intelligence was the ability to appear so) to mean that breadth itself was constitutive of intelligence – for example that an intelligent personal could comment on both cooking and Shakespeare, and that only an intelligent computer would have the breadth to replicate this performance. They then proceeded to blast holes in the formulation as relying on specific cultural capital. Yet Turing commented that computer intelligences might write great novels, but they would be written for other computers. This seems to indicate his grasp on the contingent positions of cultural perspective, and his according the artificial intelligences of the future the right to their own culture – one which would be not only non-British but in some ways entirely post-human. [39]
Rather than assessing the intelligence of the system, critics charge that the Turing Test assesses the interrogator, and its results are more reflective of the particular prejudices and preconceptions of the interrogator than of the capacities or inner operations of the system. [40] As evidence they point to the tendency of humans to anthropomorphize their technologies, the general process of confirmation bias [41] , and a more specific instance of that bias termed the ELIZA Effect. [42]
Does the Test in fact work this way?
The answer of course is yes – but this is not surprising, as Turing’s Test does not register the presence of an intelligence process, [43] but the presence of an intelligence effect. To understand the point, look at his comparison metaphor for the test. He gives the example of a terminal operator speaking to two candidates for ‘female’ – a male attempting to pass and a female control or confederate. Turing argues that intelligence could allow the man to easily pass the test – that is, ‘pass’ as a woman through the interface of the terminal. Yet Turing does not argue that this transforms the man bodily into a woman, any more than it transforms a passing computer program into a human intelligence. Rather the act of ‘passing’ has registered a significant intelligence effect – the reception of gender or intelligence by the interrogator-judge, in accordance with entirely personal and arbitrary criteria. The test itself says nothing about the producer and everything about the product – the example is more telling of the received construct of gender than of the imaginative transgendering ability of the male foil.
I return here to my central theme, that Turing should be taken at his word. “Intelligence” as it is understood in the Test is an appearance, an occurrence, a perception – specifically, an idea in the mind of the ‘interrogator’ which is produced upon consuming the productions of the ‘candidate.’
Or, in the terms of this project: Authorial identity is an effect produced by the reader examining the text…
…and any machine capable of producing this effect is as “authorial” as a human writer.
It is important to understand the complex political debates surrounding the Turing Test to understand their place in the history and the current program of machine writing and machine writing debates – however it is important not be become mired in them. [44]
By both sidestepping and reframing the issue in terms of not just statements but products, we are able to achieve the kind of distance that Turing desired for his operators using terminals – a distance that stops short of driving to the logical extreme of an Ultimate Test in which we put not just “intelligence” but the entire constitution of “humanity” on trial. [45]
While Narayan rejects subsequent discussion of Turing’s essay as productive in the present day, claiming that what has happened in discussing the game since has muddled it, he also notes that Turing never makes strict equivalence between “Can machines think?” and “What will happen if a machine takes the part of A in this [imitation] game?” Narayan further argues that Turing was laid out his game and stated his challenge of 50 years, not because he believed that the game was a strict test of thinking or that strictly thinking machines would be present in 50 years, but precisely because it would incite the kinds of activities which would lead eventually to thinking systems – and that both faith and certain goals were critical to this process. (Narayan, 17-25) The test itself was a means, not an end.
Fenstad concurs that the ends of research rather than the means of debate in the field are the thing, and wishes to sidestep the confusion generated by the Test and its philosophical debates entirely:
“The Turing test has had an important impact on the new science, giving a sense and a direction to much important research. It has moreover generated an untold number of general and philosophical papers arguing the pros and cons of the thesis. I shall not enter into this debate here. It has been useful as a check on the “naive” researcher – and AI has known several of those, even among the leading practitioners. But the general debate can be carried too far and become counterproductive. Sometimes one has to try to do things, even in the face of a priori “impossibility” arguments. And sometimes one succeeds contra the philosophers.” (Fenstad, Jens Erik. “Language and Computations.” The Universal Turing Machine Anthology. 327-347) (Yet while he makes this move, he rejects this move in the project itself – perhaps because his true goal is in itself the generation of a process rather than a product.)
Turing’s “imitation game” was a metaphor used in service of an argument that any agent capable of creating the impression of intelligence was in possession of the actuality of intelligence. His proposal was that when measuring machine intelligence, intelligence was an effect or product and not a series of causes or process, especially not some specific process that made up an underlying, authorial “real.” [46]
This is not to say that Turing was uninterested in scientific investigation or evidence – in fact, as the inventor of the “ban,” (a unit of evidence that made a hypothesis ten time as likely as it appeared before, positing the deciban as the smallest unit of evidence perceptible to human intuition) Turing lay the groundwork for sequential analysis, in which a fixed weight of evidence rather than a fixed number of experiments was required.
My stake in tying his theory to machine literature hinges on the example he chose for creating this impression – the manipulation of human language.
The extent to which interactive text itself is “literature” is its own debate, and I won’t go into hypertext or hyperliterature in depth here. Rather I’ll extend Turing’s “conversation” to literature through a simple epistolary argument before moving to discuss the effects of the Turing Test on the state and direction of machine writing today.
Give that a machine system capable of passing the Turing Test is able of commenting on a series of statements and responding to a series of questions (in conversational context), it is not too much of a stretch to imagine that the system could parse a letter into a series of statements and questions and then structure a conversational reply – there is no particular reason to prevent it from answering a page of conversation with a page, rather than a line with a line. Also, given that it could generate convincing comments and questions, there is no reason to prevent it from processing its own outputs – that is, generating a question, and then responding to its own question, and so on. This is to say that a system capable of passing the Turing Test should be able to write rudimentary dialogs. While some might argue that a single line response to a typed question or statement is not truly “writing” or “literature,” this is a technicality of temporal arrangement – and these hypothesized letters or dialogs generated by the same system are perhaps not so far removed from the art form of the essay or the short play. This argument is crude, but it will have to stand for now as we go on to discuss what Turing machine-writing systems are available for study in the present day.
Machine / mechanical / mathematical writing goes back centuries to early religious texts, and especially in permutational or combinational poetry, and its various forms depend on the breadth of your definition of machine or math as well as your definition of writing (consider for example some of the formal structures of musical composition).
In this context, I’m interested in machine writing as a) an externalization, and as such a system, which once designed, writes semi-autonomously, and b) a field influenced by Turing, and thus its developments in computer science since the 1940s.
One side effect of Turing’s test theory was to encourage generations of computer scientists to take up the challenge. By drawing a line in the sand, Turing encouraged them to write programs specifically geared towards passing that line – by hook or by crook. Thus programs like ELIZA and its descendants have functioned not by modeling though and then learning to express it convincing dialog, but by modeling convincing dialog with little or minimal ‘thought.’
Instead of outlining a modern examples in detail, here I’ll instead sketch and refer interested parties to my Constrained Authorings site, co-authored with Elizabeth Freudenthal, and our notes for a Machine Literatures lecture. [27]
Many of the programs on the web currently used for machine writing accomplish their products through substitution (reference and filtering) and rearrangement (anagramming, permuting, and cut-up).
Substitution systems tend to be highly simplistic and generally used for humorous effect. One example is sites that generate fanciful names for people based on feature characteristics of their real names (Rock Name, Glam Name, Secret Agent Name, Porn Name etc). Filtering machines generally rewrite content in a particular dialect or idiom, such as a machine baby-talk filter or old-testament filter used to view other web pages. The techniques of these sites are often limited to simple word-by-word substitution. On the high end of filtering machines are the much more complex in-line translators, which actually run the entire text through a language pre-processor/compiler before attempting to rewrite the entire text into another languages – generally poorly, but in a way that might lend at least a minimum level of legibility.
In contrast to substitution systems, rearrangement systems manipulate the linearity of existing texts. A great example is Here Comes Everybody, a site that reprocesses Finnegans Wake into pages synthesized from fragments taken from all over the text.
Part of the success of Here Comes Everybody is the design of the system with respect to a particular (and friendly) text subject. When combined, substitution and rearrangement rules in conjunction with a specifically designed or selected text-resource have produced some of the most “literary” machine writings – haiku produced from dictionaries, being one of the most common examples.
However these systems, while more sometimes more effective in limited contexts, remain far less complex than the conversation bot programs (called “Chatterbots”) that have been the primary attempts to pass the Turing Test.
“A bot is a software tool for digging through data. You give a bot directions and it brings back answers. The term bot has become interchangeable with agent, to indicate that the software can be sent out on a mission, or a crawl, or data mining. Bots were not invented on the Internet, however. Robotic software is generally believed to have been created in the form of Eliza, one of the first public displays of artificial intelligence....”
(http://www.botspot.com/bot/what_is_a_bot.html)
While the editors of BotSpot.com err in calling Eliza an “artificial intelligence,” they do identify a key link between bots-as-agents and conversation chatterbots of the Eliza type. Like many programs which would follow, Eliza was a black-box program designed which added a little sophistication to an otherwise simple set of substitution and rearrangement rules – when interrogated, the program would cut up and respond to the interrogation after the manner of a psychoanalyst, constantly presenting the interrogator with a linguistic mirror.
Projects since Eliza have gained much in the way of sophistication (CoLIN is a recent example), [28] however they have maintained much of their strategic bent towards mirroring. This in and of itself is fascinating – the terms of the Turing Test seem to presuppose a reader-centered rather than writer-centered mode of address in order to have a chance of success, and this mode has become constitutive of the tightening cycle of chatterbot production and chatterbot evaluation.
To one side of this process continues the attempt to make computers think, and then speak. Such projects include ThoughtTreasure, a logical articulation system that produces “common sense” statements and assertions drawing on a database of “thoughts.” Although ThoughtTreasure currently must speak and be spoken to in an intermediate programming language (rather than a natural one, which might “pass” for human), it is an example of an attempt to build the back-end for a kind of bot-agent for which a chatterbot might provide the front end. [29]
This indicates that the database-driven revolution in information systems may bring about a re-merging of Turing’s broad vision for the integrated machine agent. Narayanan (16) Marks the beginning of the fields of Computational Philosophy and Philosophy of Artificial Intelligence with Turing’s 1950 “Computer machinery and intelligence’ paper. He notes that computational theorists are most influenced by his ‘Turing Machine’, AI workers by his ‘Turing Test.’ Narayanan cites Chomsky as the bridge between the formal properties of grammars and automata, connecting many of the fruits of Turing’s investigations to the textual matter he eventually hoped they would address. [30]
ThoughtTreasure is also a compelling example due to its built in reconciliation of error and its and “best-guess” processes when confronted with data which is inherently contradictory. Brosl Hasslacher argues in the article “Beyond the Turing Machine” (The Universal Turing Machine anthology p417-433) that complexity studies had revealed problems that could not be addressed by Turing Machine style architectures, and required a new conception of computers and programs with chaotic systems integrated into their operations. While ThoughtTreasaure is far from this radical paradigm shift, it does begin to address these topics.
The future of substitution writings may be systems after the style of the Kurzweil Poet [31] – partial, hybrid, and interlocking with informed and consenting users who do not interrogate so much as they read. The future of rearrangement writings might lie in the realm of dynamic logical databases [32] which reformulate questions to think through answers – or in database driven cutups which are the precursors of self-organizing text systems (a shameless plug for reading my Self-Organizing Maps project). [33] So long as we continue to approach the production of text in new ways, the spirit of the Test reminds us that we must both evaluate our system productions aesthetically and take the results of those evaluations seriously, for the have independent reality.
The biggest mistake we could make at this point is to pursue the superficial details of the test to their logical extreme in examinations such as The Ultimate Turing Test. This cyber-inquisition endows a Turing subject with all of the characteristics of a human being in the world, and then examines them for jerky motion, unnatural mannerisms, odd pacing or tone of language, etc. It is both the ultimate fantasy of the robot/program “passing” for human in an all too physical cyberspace, and re-imagines Turing’s investment in the autonomy of machine intelligences and productions as a maniacal program to recast them in our image. Ironically, Turing would probably have failed The Ultimate Turing Test – his nervous habits and awkward mannerisms of social speech would doubtless have given him away as some form of cyber-replicant
Although Turing imagined computers one day writing novels for the enjoyment of each other, in complete disregard for human aesthetic needs, he did also credit them with the capacity for creating in humans the narrative response – the sense that there was someone behind the words. This evocation of voice did not need to be synonymous with the reality of an intelligent presence – any more than the strong reaction of London readers to Sherlock Holmes had signaled the existence of a real doctor Watson. Or for that matter any more than the reaction of some readers to ELIZA had signaled the presence of a real psychotherapist. Part of the magic of the reader for Turing– both wonderful and horrible – was the ability to experience something with as it was to them. Thus the reality of the computer code and the reality of the reader’s relationship to ‘the author’ could coexist meaningfully. This belief not just in multiple identities but in the reality of identities projected on others may have been what led to his full confession and his plea of no contest in court, combined with his continual insistence that he had done nothing wrong. For Turing, a person could be both who he was and who people saw him as simultaneously – and in fact had no choice. Although Turing lived uneasily with the ‘imitation games’ and intermediate processes of his own life, he simultaneously wondered if there was in fact anything else.
“In considering the functions of the mind or the brain we find certain operations which we can explain in purely mechanical terms. This we say does not correspond to the real mind: it is a sort of skin which we must strip off if we are to find the real mind. But then in what remains we find a further skin to be stripped off, and so on. Proceeding in this way do we ever come to the ‘real’ mind, or do we eventually come to the skin which has nothing in it?” - Turing [34]
Today, Turing’s program for is running a little behind schedule. [47] The bots that are Turing’s legacy are evolving year by year, producing writings on demand which attempt to ‘pass’ – sometimes as female, sometimes as male, always as human. Although they are not intelligent even by Turing’s standards (the challenge of the Test has yet to be met) these limited systems are producing texts, and these texts are beginning to pass with humans as human-authored. At the same time, researchers in neural networks and natural language are developing systems that cannot yet speak or write with any great coherence, but can begin to learn.
Perhaps these projects must mature before any truly literary machine writers can emerge, and the products of those writers will display have merit in direct proportion to the complexity of their underlying “thoughts.” However perhaps writing systems of great sophistication can precede thinking systems. Could great works might come from authors who speak a language, but don’t think with it... or perhaps not. If the critics of “black box trickery” are right, then we must take an enormous step backwards in articulation from Chatterbots before we take the next small step forward in machine writing, because we will need AI – and attempting to narrowly pass the Turing Test is not AI.
However it is important to remember that easily half of the language apparatus needed to make machine texts readable is already present in the person of the reader. Once the text is read, and found to be good, the system that produced it may be as much of a black box to us as the mind of a distant author is to today’s reader. Then we sill continue the great 20th century argument over the importance of understanding the author to understanding the work – the same discussion with a high-tech face. While I am not particularly interested in reinstating the New Criticism divide between text and author, I am invested in expanding our definition of the literary work, uncovering and accepting the long codependence of language and language technologies, and strengthening the connection between the text and the reader. Other critics with other stakes will enter the arena, and I hope to participate in preparing the way for the great debate.
Whenever and however we reach the moment of machine authorship that does ‘pass’, we will have the guidelines and the philosophical preparation to understand them as human works – human not because the arrangements of text were assembled by a human, (which they were not) but human in the same way that the works of Shakespeare are human documents when (and in a way how) I read them, and would be so even if my copy had somehow been generated by a monkey dancing on a typewriter.
Or on a computer.
Aarseth, Espen J. Cybertext: Perspectives on ergodic literature. Baltimore: Johns Hopkins UP, 1997.
Crosson, Frederick J., and Sayre, Kenneth M., eds. Philosophy and Cybernetics: Essays delivered to the philosophic institute for artificial intelligence at the University of Notre Dame. Notre Dame: Notre Dame UP, 1967.
Gillies, Donald. Artificial Intelligence and Scientific Method. New York: Oxford UP, 1996.
Herken, Rolf, ed. The Universal Turing Machine: A half-century survey. Oxford: Oxford UP, 1988.
Hodges, Andres. Alan Turing: The enigma. Essex: Burnett Books, 1983.
Landow, Goerge P., ed. Hyper / Text / Theory. Baltimore: Johns Hopkins UP, 1994.
McCorduck, Pamela. Machines Who Think: A personal inquiry into the history and prospects of artificial intelligence. San Francisco: W.H. Freeman, 1979.
Murray, Janet H. Hamlet on the Holodeck: The future of narrative in cyberspace. New York: Free Press, 1997.
Mishkoff, Henry C. Understanding Artificial Intelligence. Sams understanding series. Indianapolis: Macmillan, 1985.
Nayrayanan, Ajit. On Being a Machine: Formal aspects of artificial intelligence. Vol. 1. Ellis Horwood series in artificial intelligence foundations and concepts. Halsted Press: New York,1988. 2 vols.
Thornton, Christopher James. Truth From Trash: How learning makes sense. Complex adaptive systems. Cambridge Mass: MIT Press, 2000.
Turing, A.M. Pure Mathematics: Collected works of A.M. Turing. Ed. J.L. Britton. Amsterdam: North Holland, 1992. 207-210.
Turing, Sara. Alan M. Turing. Cambridge: W. Heffer & Sons, 1959.
Young, J.F. Cybernetics. London: Illife Books, 1969.
The Turing Test Page. http://cogsci.ucsd.edu/~asaygin/tt/ttest.html.
The Turing Test and Chinese Room Experiment. http://members.aol.com/lshauser/turingho.html.
Alan Turing, In Brief. http://www.turing.org.uk/turing/.
Kulturezone. http://www.evolutionzone.com/kulturezone/index.html.
P3RMUT4TI0NS. http://userpage.fu-berlin.de/~cantsin/index.cgi.
Cent mille milliards de poèmes. http://userpage.fu-berlin.de/~cantsin/queneau/poemes/poemes.html.
VISPO - Langu(im)age Poetry - New Media: Links of the Imagination. http://www.vispo.com/misc/links.htm.
A Fairy Tale as You Like It. http://userpage.fu-berlin.de/~cantsin/queneau/conte/conte.cgi.
Cut-ups and the Internet. http://www.dsl.org/comp/cutups.shtml.
hAIku. http://www.obs-us.com/people/sunny/haiku/haiku.htm.
Random Word Haiku http://www.cs.indiana.edu/cgi-bin/haiku.
Here Comes Everybody. http://userpage.fu-berlin.de/~cantsin/n-8/aleph.cgi?&q=river&i=w.
The Neoism Machine. http://userpage.fu-berlin.de/~cantsin/neoism/cantsin.cgi.
BotSpot. http://www.botspot.com/.
CoLIN (ComputerLinguistics ImitatioN). Brown, Alan J.
The Age of Intelligent Machines: “A (Kind of) Turing Test.” Kurzweil, Ray. http://www.kurzweilcyberart.com/poetry/rkcp_akindofturingtest.php3.
Wired Magazine. Idées Fortes. The Lying Game. Wallace, Richard S. http://www.wirednews.com/wired/archive//5.08/idees_fortes.html.
Twelve Reasons to Toss the Turing Test. http://zompist.com/turing.html.
The Ultimate Turing Test. Barberi, David. http://www.ibiblio.org/dbarberi/vr/ultimate-turing/.
In this essay, I’ve begun to frame the issues and the theory I wish to bring to bear on the history of machine writing. The second half of this work will deal with applying this theory to the reality of that history, both in terms of the specific programs and contests, and in terms of a general picture of the emergent technology which today owes its genesis to Turing’s theories. Understanding theoretical and philosophical underpinnings of our current drive to produce machine texts is vital to evaluating the terms of the machine authorship debate – and I am convinced that machine authorship will be key understanding the articulated internet in which our data becomes discourse.
I refer to this idea in my Self-Organizing Maps guide. This paper is part of an ongoing digital humanities project, which will link my writings on Turing, Self-Organizing Maps, Intelligence and Genius, Machine Writing, Postmodern Identity and Queer Theory, and Hypertext Narrative into a loose constellation of critical nodes.
Major sections currently include:
[1] This could be algorithmic in a variety of senses – combination, permutation, substitution, or much more complex systems up to and including freeform grammatical composition of sentences and passages, either haphazardly or in response to some question or topical stimulus.
[2] One example of a non-computer writing machine might be a set of rules written down on a sheet of paper for creating a Mad-Libs poem - so long as the words were supplied in accordance with the rules and using some predefined process such as rolling dice to select a page from a dictionary and a word from that page. The directions page, the dictionary resource, and the dice operation taken together constitute the machine.
[3] “Computing Machinery and Intelligence.” Mind, Vol.59, No.236, pp.433-460. Complete text: (http://www.oxy.edu/departments/cog-sci/courses/1998/cs101/texts/Computing-machinery.html).
[4] Sometimes called a confederate, player, or foil.
[5] For most purposes, the judge and the interrogator are considered to be the same terminal operator.
[6] Test diagram (http://www.macrovu.com/CCTMap2TuringRm.html).
[7] The Turing Test Page (http://cogsci.ucsd.edu/%7Easaygin/tt/ttest.html) is a comprehensive online resource on the Turing Test (not the man himself) and issues surrounding it, with great links.
[8] Sidelined is the entire question of whether human intelligence is a unified entity (rather than intelligences), or indeed if our socially constructed concept of intelligence truly refers to the inner processes of thought (rather than the results of those processes). While this paper could extend its arguments of art and reception from against author to against intelligence (or genius) as such, these arguments will not be made here.
[9] For more on this see such resources as www.BotSpot.com.
[10] An ongoing competition to find the first computer program which can pass an unrestricted Turing test.
[11] Turing stated that the truly literate and intelligent machine might write great novels – but it would presumably write novels which were aesthetically appealing and of interest to other computers, rather than to humans. This frames even more sharply the projects of machine writing as product rather than process oriented – for in order to evaluate a legitimate process of machine writing as such, rather than passing as human writing, Turing points out we require a panel of machine readers to render judgment.
[12] For example, a computer can iteratively display the results of highly complex calculations faster than the human eye can refresh images and present them to the human brain – let alone have the brain process or comprehend those results, or the calculations behind them.
[13] Gillies, 114-121. The Anxieties Caused by Advances in Artificial Intelligence. (This includes a wonderful quote on Penrose).
[14] Narayanan, 38. “The modern counterpart is the view that AI is dehumanizing because of its adoption of a certain view concerning what it is to be human, viz. that a human being and his or her behavior can be described or explained using the terminology and concepts of computational theory. Modern objectors argue that AI is yet one more attack from scientific and philosophical quarters on the notion of humanity. We shall not follow this up in these two volumes: we only point out there is a significant portion of AI literature devoted to this subject (see for instance Boden (1981), Weizenbaum (1976), Rogers (1984)) and if Turing were to make a similar claim now, he probably would not be allowed to escape so easily.
[15] Turing’s 1950 essay included his parody of the “Heads in the Sand” Objection:
“The consequences of machines thinking would be too dreadful. Let us hope and believe that they cannot do so. This argument is seldom expressed quite so openly as in the form above. But it affects most of us who think about it at all. We like to believe that Man is in some subtle way superior to the rest of creation. It is best if he can be shown to be necessarily superior, for then there is no danger of him losing his commanding position…. It [the popularity of the argument] is likely to be quite strong in intellectual people, since they value the power of thinking more highly than others, and are more inclined to base their belief in the superiority of Man on this power.”(15)
[16] For comparison, read the censored view presented by his mother in Sara Turing’s Alan M. Turing. Online, try the site “Alan Turing, In Brief” at http://www.turing.org.uk/turing/. For primary sources, consult J.L. Britton’s Pure Mathematics: Collected works of A.M. Turing.
[17] The Test, which Hodges (415) termed ‘a sexual guessing-game’ … “Although pleasantly recalling the secret messages that might be passed in his conversations with Robin and Nick Furbank, this was in fact a red herring, and one of the few pages of the paper that was not expressed with perfect lucidity. The whole point of this game was that a successful imitation of a woman’s responses by a man would not prove anything. Gender depended on facts which were not reducible to sequences of symbols. In contrast, he wished to argue that such an imitation principle did apply to ‘thinking’ or ‘intelligence’. If a computer, on the basis of its written replies to questions, could not be distinguished from a human respondent, then ‘fair play’ would oblige one to say that it must be ‘thinking’.
[18] On February 3, 1952, Turing was interrogated by two detectives and arrested for homosexuality.
[19] Hodges, 460. “The conservative fear was that every kind of behaviour would be excused by appeal to some irresistible, uncontrollable force majeure… they sought a non plus ultra to pretensions of mental determinism, a barrier against the flood of threats to traditional values unleashed by the Second World War. They found one in homosexuality: The new men’s talk of ‘conditions’ and ‘complexes’ was not to be allowed to excuse a deadly social evil, corrupting and weakening everything in its path
[20] Hodges, 468. “In 1944 Dr. Glass attempted to decrease the sex drive of homosexual men by injecting them with male hormones. However many of the subjects reported an increase in the homosexual drive after treatment. So science turned to the work of C.W. Dunn, who in 1940 had experimented with estrogen – and found that it did reduce the sex drive for heterosexual and homosexual men alike.”
[21] Hodges, 173-174. “In September, 1938, when Germany changed the operation of its Enigma machine (to change initial settings and self-encode them), the only remaining way of establishing a fingerprint was the repeated triplet – a letter occurring over again which indicated that it encoded to itself. Thus AGHTUITUI would encode to AGHRYNFYPA, with the Y Y being a repeated triplet, called a ‘female.’” (diagram 180)
[22] Hodges, 239. “As chief consultant to GC and CS, he was living at the heart of yet another imitation game, doing work which did not officially exist. Now there was almost nothing in his life that he could talk about but chess playing and fir cones.”
[23] Hodges, 474. “His friends argued for scientific ‘organotherapy’ rather than prison, little realizing its effects… The treatment caused clinical depression and also the growth of breasts.”
[24] Hodges, 489. Turing’s mother’s interpretation, and Hodges theory of it as the perfect crime.
[25] : Hodges, 527. “According to his imitation principle, it is quite meaningless to speculate upon his unspoken thoughts… but Alan Turing could not possess the philosopher’s detachment from life. It was, as the computer might put it, the unspeakable that left him speechless.”
[26] Sara Turing’s 1959 biography was also silent on Alan’s sexuality– the subject would not be reopened until 1970s.
[28] Computer Linguistics ImitatioN by Alan J. Brown. A free ChatterBot program based on the premise by Alan Turing, the computer pioneer who decreed that Artificial Intelligence would exist if a computer could make a human believe that they were talking to another human. Most ChatterBots pretend to give human responses, CoLIN actually learns language from scratch, as a child would. This unique approach also gives CoLIN the ability to hold a conversation in any language. CoLIN is designed to respond to user statements based on previous statements made by the user. In other words, it learns what to say by listening to you. http://www.barc0de.demon.co.uk/nlp/
[29] Under development since 1994, ThoughtTreasure is a comprehensive platform for natural language processing (English and French) and commonsense reasoning. ThoughtTreasure contains pieces of common sense…. (from overview http://www.signiform.com/tt/htm/overview.htm)
[30] Fenstad. Using Chomsky’s 1957 Syntactic Structures work, Fenstad worked on schemas for translating natural language questions into grammatical structures which could be operated on with respect to a knowledge base, then translated back into natural language answers. In attempting to approach the Turing test “in a principled way; we are not interested in a ‘black box’ filled with ad hoc trickery”(346) Fenstad set himself an enormous task, however began what may be the distant second wave of machine writings – agents which think and then speak rather than manipulating speech as a mode of thought.
[31] The Age of Intelligent Machines: "A (Kind of) Turing Test." http://www.kurzweilcyberart.com/poetry/rkcp_akindofturingtest.php3
[32] : Thornton, 73-90. “What this tells us is that the Bletchley Park operation and the Polish work (on the German Enigma during WWII) that proceded it constitute the first and perhaps most dramatic examples of large-scale computational learning. But there is something more… it involves identifying and disentangling the relationships that exist in the data, utilizing available knowledge about the processes that generated those effects.”
[34] Also see Oswald Wiener “Form and Content in Thinking Turing Machines” (631-657) A meditation on whether or not the brain is truly onion, layers which may be all restated in terms of Turing Machines. At stake is whether we can hope to understand ourselves, or if it is beyond our capacity. Seems to respond to Searle’s 1980 criticisms (Chinese Room) and his notion of “causal powers of the brain” which presumably cannot be explained. Reaction against. Sees algorithmic understanding as sufficient, only problems remain in ‘implementation’ – complexity one of those implementation issues.
“A wholly different question, of course, is possible criticism under the image of the self as still entertained by the humanities. I believe that here radical reorientation is inevitable. Introspection in the purpose of establishing spontaneity and irregularity of the mind is the main source of the anti-mechanist; but this is a far cry from observation intended to track recurrence of events, to assist in generalization and heuristic modeling. Like every other kind of observation, sound introspection will never tell what things are, but just what they do in terms of performed notions. Therefore consciousness will have to be described in terms of process: it is what makes introspected things behave the way they do.”(655)
[35] : Gillies, 151. “Why Advances in Computing are more Likely to Stimulate Human Thinking than to Render it Superfluous.” A reconciliation of the Lovelace-Godel incompleteness objection with a belief in Turing-Penrose predictions – machines will exceed in all areas, but humans will remain politically superior, as well as always possessing a super-set of information in comparison to the machine set of information.
[36] Turing’s typically rejects the preoccupation with means rather than ends as ludicrous: “We could produce fairly accurate electrical models to copy the behavior of nerves, but there seems very little point in doing so. It would be rather like putting a lot of work into cars which walked on legs instead of continuing to use wheels.”
[37] Narayanan, 37-43. The remained of the text is devoted to teasing out specific objections to the Test’s assertions of machine-thought and developments in them over the intervening four decades.
Some of the restated and refined modern objections in the field of AI are summarized at “Twelve reasons to toss the Turing Test.” http://zompist.com/turing.html
[38] A discussion of The Chinese Room experiment, its analysis and objections is called for here – hopefully to be elaborated at a later time. Searle’s critique is actually quiet complex and sophisticated, and also shook up the field of artificial intelligence and artificial neural networks for years – it deserves a fair hearing. The Turing Test and the Chinese Room Experiment (http://members.aol.com/lshauser/turingho.html) contains a tabled writeup with diagrams and references.
[39] : Narayan asserts that Turing’s move to the imitation game was a sidestep around Descartes and the axioms of dualism. He also argues against an interpretation of Turing’s move as behaviorist, as behaviorism would create a ‘separate but equal’ standard around the term intelligent, while Turing thought machines could be intelligent on human terms.
Fenstad, Sections 8-10: Communicating Concepts and Communication Protocols, Cultural Systems, Computing vs. Intelligence (p46-463) Discusses the dissemination of CS and AI ideas in a cultural context, and the evolution of human-machine communication as a socializing process forming over time.
[40] Determine Whether Computers Can Think?: The History and Status of the Debate (http://www.macrovu.com/CCTMap2.html) contains a graphical representation of the main debate points.
[41] A psychological process whereby humans interpret differences they encounter by incorporating them into the patterns of their expectations
[42] “ELIZA effect” coined by Douglas Hofstadter, 1995. The tendency of humans to read more into computer performance than is warranted by their underlying code. First described by Joseph Weizenbaum, creator of ELIZA
[43] As Johann A Makowsky notes in his article “Mental Images and the Architecture of Concepts,”(453-465) Turing’s models of computability refer to output, but “are inadequate to describe the manipulation and transfer of concepts without explicit reference to the coding of these concepts.”(454).
[44] : Narayan, 18. Rejects subsequent discussion of Turing’s essay,
[45] The Ultimate Turing Test “Can Turing’s test be improved on? Yes. With current advances in computer graphics, virtual reality, biomechanics and many other fields, it is possible to create an "Enhanced" or "Virtual" Turing test. The underlying idea of the test is still the same, but the amount of interaction between judge and subject is increased greatly….” http://www.ibiblio.org/dbarberi/vr/ultimate-turing/
[46] : Hodges, 266. “Implicit in these discussions was the materialist view that there was no autonomous ‘mind’ or ‘soul’ which used the mechanism of the brain. (He had perhaps hardened his stance as an atheist, and his conversation was more free with anti-God and anti-church jokes that it would have been before the war.)”
While he discounts complexity theory, Roger Penrose “On the Physics and Mathematics of Thought” (491-522) notes that Turing seemed to consider the human brain a subset of the universal Turing machine, with finite storage capacity and temporal range, as well as the capacity for mistakes. He argues that brains are ‘a physical device’, and asserts that as their operations involve not only chemistry and electrodynamics but quantum processes, their operations are essentially non-algorithmic in nature.
[47] “Turing predicted that within 50 years (by the year 2000) technological progress would produce computing machines with a capacity of 10^9 bits, and that with such machinery, a computer program would be able to fool the average questioner for 5 minutes about 70% of the time.” (The Free On-line Dictionary of Computing. 1995. Updated 9-5-2000)
Text and Test | Against Machine Writing | Against Queer Identity | Against Turing Tests | From AI Impersonation to Agents | (un)Conclusions