May 31, 2008

Provocation by Program: Imagining a Next-Revolution Eliza

by Nick Montfort · , 12:17 pm

By Nick Montfort and Andrew Stern

(This is the text of the talk we gave at the ELO Visionary Landscapes conference just now. Mark Marino already has a reply online.)


In the 1960s, Eliza, and specifically that system running the famous Doctor script to impersonate a psychotherapist, prompted conversations and controversies about anthropomorphic concepts of the computer, artificial intelligence and natural language understanding, user interfaces, and even psychotherapy. Decades later, Janet Murray hailed the system as the first electronic literature work, saying it was at that point still the most important one. All this was the result of a rather small amount of code that lacked multimedia elements, contained very little pre-written text, and was developed by a single person, Joseph Weizenbaum.

We begin by assuming that computation and literary art are inherently very powerful. That is, we assume it is not essential to have recourse to networked communication, massive knowledge bases, or even graphics capabilities to develop a provocative, affecting project that inquires about important issues. In thinking about a such a project, we are seeking an antidote to today’s ever larger and complex computer applications — sixty-hour game quests within expansive virtual worlds, mashups of intricate Web technologies, and massively feature-bloated operating systems. A small yet powerful and surprising computer program would be both pleasurable and provocative because of its simplicity and clean concept. So we simply assume, rather than trying to prove, that while more elaborate systems may be interesting in some ways, a new system on the scale of Eliza can still have the sort of broad impact today that Weizenbaum’s computer character did more than forty years ago. Given that, we ask, what specific qualities would this system have?

Of course, there are plenty of programs that are more or less directly descended from Eliza. These include conversational characters (such as A.L.I.C.E.) as well as task-oriented dialogue systems (such the automated Amtrak agent Julie that you can call and speak to right now at 1-800-USA-RAIL). There are also digital artworks that specifically refer to and rework Weizenbaum’s concept, such as Adrianne Wortzel’s Eliza Redux. We are not imagining programs along these lines as we think about a possible Eliza for the 2000s or 2010s. Instead, we will focus on the qualities that made Eliza a provocative and influential piece of literary art in its time and context. We are not thinking of a new chatterbot; we are interested in imagining a system that would introduce a new form, like that of the chatterbot, and that would inspire reworking and reimagining by artists, as Eliza did.

Six Important Aspects of Eliza

While different writers, artists, and programmers might identify different aspects of Eliza as central, we believe that we have identified six that would be shared by a similarly high-impact program today. Some of these may be fairly obvious; others seem to us to much much less evident and much less frequently discussed, if they have been discussed at all. The important properties of Eliza that we have identified are:

  1. Engaging deeply with language.
  2. Dealing with a fundamental issue, concern, anxiety, or question about computing and technology.
  3. Being interactive and immediate — impressing the interactor in an instant.
  4. Being understandable after more effort is applied and the program is explored further.
  5. Being general to different computer platforms and easily ported.
  6. Being process-intensive — driven by computation rather than data.

Next, we discuss each of these aspects of Eliza in a bit more detail, explaining why we think each of these was important to Eliza’s effect:

Other Important Systems and the Qualities They Share

We now look to several other computer systems, small-scale and large-scale, that have become part of the zeitgeist, and we consider which of these qualities they have and which they lack. These high-impact systems are not restricted to literary ones; in fact, we did not select any system that was mainly framed as literary.

Of course, these systems had some important characteristics that Eliza didn’t — in certain cases they could be sold as computer games, or connected to advertising to form the basis of a profitable business, and so on. It does seem, however, that elements of their success and influence were shared in common with Eliza. If we were looking to stage a Cold War demonstration or deploy a hit video game, we might consider a different set of attributes, but we believe Eliza’s qualities are particularly worth considering for those of us who are involved with digital writing, with imaginative and poetic uses of language.

Directions for Impactful E-Lit

To conclude, we will identify the types of electronic literary practice that we believe most likely to have Eliza-like impact. All sorts of literary practices are well situated to engage language, and many practices also deal with questions about computing technologies and their situation in society. The remaining four aspects are not shared equally by works coming from different electronic literature practices.

Electronic literature authors often try to invite deep reading and to reward lengthy exploration; perhaps because the involvement of a committed reader is valued over an initial glance, impressing the interactor right away is seldom an important goal. But literary art on the computer can make an immediate impression: While the hypertexts of the Eastgate school offer a great deal to the careful reader, William Gillespie, Scott Rettberg, and Dirk Stratton’s The Unknown grabs and slaps the reader more or less immediately with its humor and metafiction, no matter what page is selected.

In terms of yielding to understanding, some of the art and literary works that are initially the least understandable actually display this quality the most clearly. Jodi’s maze of broken websites can increasingly be seen to display the logic of the Web; Ben Benjamin’s Superbad can be understood as an ecstasy of mid-1990s Web design; the malfunctions of Dan Shiovitz’s Bad Machine give information about the virtual situation as the interactor learns more about them.

Generality and portability are very seldom valued in electronic literature practice, but there are some signs that this is changing. Rob Kendall’s X-Literature prototype is an important step toward abstracting the functioning of hypertext-like works, and, aside from offering preservation benefits, should urge electronic literature authors of all sorts to think about the essentials of how their works function.

The least process intensive electronic literature works are often held up as paradigms, but Eliza is not alone in being computational and working in literary ways. With regard to emphasizing computation over data, computational poetry, interactive fiction, interactive drama, and creative text generation practices are more process intensive, more Eliza-like, and most likely to connect computing and culture in the way that Weizenbaum’s program did.

Of course, many electronic literature authors seek to express an imaginative world or poetic concept in other ways, without modeling aspects of language in an interactive program. Eliza-scale provocation is not a goal shared by everyone. But we hope that our analysis of Eliza, performed from our standpoint as computer literary artists, is nevertheless of interest to certain provocateurs.

27 Responses to “Provocation by Program: Imagining a Next-Revolution Eliza”

  1. Andrew Stern and Nick Montfort have a provocation from ELIZA at WRT: Writer Response Theory Says:

    [...] “Provocation by Program” Imagining a Next-Revolution ELIZA Nick Montfort and Andrew Stern [...]

  2. Mark Bernstein Says:

    > While the hypertexts of the Eastgate school offer a great deal to the careful reader, William Gillespie, Scott Rettberg,
    > and Dirk Stratton’s The Unknown grabs and slaps the reader more or less immediately with its humor and metafiction,
    > no matter what page is selected.

    Though this is perhaps an archetypical ELO conference provocation, I wonder how far we’d want to extend the argument. Let’s face it: humor and metafiction are an odd couple! And do we really want to honor primarily, or exclusively, the grab and slap? If grabbing and slapping on every page is the aesthetic goal of The Unknown, why is The Unknown so long?

    It might also bear consideration that the interpretation of Weizenbaum’s literary work you offer in the opening section, while shared by a number of writers including Murray, is directly contrary to the interpretation Weizenbaum himself offered. I mention this not to revive the intentional fallacy (though in this specific case it might have some bearing) but because this circumstance seems as pertinent to the reception and success of ELIZA as do the others you adduce.

    Finally, might the true explanation for the ready dissemination of this program lie not so much in ease of portability and documentation as in the trivial nature of the underlying code — and the ubiquitous foundation narrative of the discursive computer, for which ELIZA users had been prepared by decades of literature?

  3. scott Says:

    I agree. Many pages of the Unknown have no grabbing, slapping, or even heavy petting. And some of them aren’t even funny. I weep nearly every time I read of Dirk’s death.

  4. Mark Marino Says:

    Mark raises a good point. Disowning our works once we’re are done might be a good move toward pushing the conversation along. I’ve been having some misgivings about the humanistic implications for some of the work we’ve been doing over at Bunk Magazine lately. The Los Wikiless Timespedia seems to have caused people to misunderstand the nature of newspapers.

    I think your last point is right on and speaks to Nick and Andrew’s second point of criteria. Of course, I can only imagine that at least some of the literary works we are discussing, from Blade Runner to Galatea 2.2 are in direct dialogue with ELIZA.

    Mark, come over to WRT when you are done here. I’ve got some assertions about Storyspace I’d like to run past you.

  5. Amir Michail Says:

    Check out the Chatbot Game, a Web 2.0 approach to building a chatbot:

  6. nick Says:

    Mashups and the network strike back.

  7. Richard Evans Says:

    Very interesting article.

    There is one other property of Eliza which was also important to its success: the simulated environment was carefully chosen to mask the weaknesses of the simulation. The Rogerian therapist (or a caricature thereof) doesn’t really need to understand what the patient is saying – he just makes a merely syntactical manipulation to turn an assertion into a question.

    Facade is another example where the simulated environment was carefully chosen to cover up the simulation’s problems. The characters are neurotic and self-absorbed; they are acutely aware of the modernist worry about a lack of meaning. These aspects of their personalities are perfect for covering up some of the cases where the parser doesn’t quite understand what you are saying!

    (NB I do not mean this as a criticism – it is part of the success of Facade that the authors anticipated in advance the inevitable limitations of the technology, and designed a dramatic situation which de-emphasized them).

  8. andrew Says:

    Hi Richard, you make great points, as usual!

    What you say is certainly true about Eliza (and hopefully Facade too, for that matter).

    In terms of this discussion, though, identifying general traits that made Eliza important and high-impact, if we were to name another high-level characteristic, how would we describe the one you mention?

    We could call it something very general like “good design” or “cleverly covers technical weaknesses”, but those are universal design maxims, the kinds of traits one would want in any production.

    I think the traits we have identified are perhaps a bit higher-level than the particular feature of Eliza you point out?

  9. Richard Evans Says:

    Um… good question… perhaps something like “The computational weaknesses of the simulation are mirrored in the psychological weaknesses of the simulated protagonist” ?

  10. Richard Evans Says:

    Another striking example of this is Emily Short’s Galatea. The NPC character is an animate statue that has only been alive for a few days. These two aspects of the fiction perfectly excuse the fact that computer implementations of natural language understanding lack emotional understanding and real-world knowledge: she lacks emotional understanding because she is a statue, and she lacks real-world knowledge because she was only born yesterday!

    In all these cases (Eliza, Facade, Galatea), the inadequacy is both highlighted and also excused. It is made explicit so that we can forget about it.

  11. Matthew Johnson Says:

    Apropos of little, did anyone catch the Eliza reference on an episode of Terminator: Sarah Connor Chronicles?

  12. nick Says:

    I didn’t, but it’s not off topic at all. If you let me in on what the reference is, and if Andrew and I do some sort of for-real publication based on this work, we might even cite this as an example of Eliza’s influence.

  13. andrew Says:

    I do watch and enjoy the show, and I too remember a reference to Eliza in one of the episodes. Part of me thought, okay, yeah, the writers are just trying to show us they’re all knowledgeable-and-shit about computing and AI. But actually you’re right, it probably is a notable sign of Eliza’s influence on today’s popular consciousness.

    Of course, one of the things I like most about Terminator: The Sarah Connor Chronicles is the fact that it has a handsome young male character who almost single-handedly invents the most powerful AI in the world in his bedroom, and happens to be named Andrew.

  14. Ian Bogost Says:

    Richard, Andrew: I’m not sure I understand how Facade does something similar to Eliza and Galatea in terms of modulating the player’s expectations of the agents’ inadequacies. In fact, isn’t one of the common critiques of the aesthetic effects of the NLP+ABL infrastructure precisely that it fails to do this?

  15. andrew Says:

    As Richard describes, it’s in Grace and Trip’s character design — their neurotic and self-absorbed nature helps players suspend disbelief when the NLP fails. (And of course we write dialog responses to try to cover up as best we can, the times the NLP knows it’s failing.)

    Sure, there is plenty of complaint that the NLP fails — but there would perhaps been even more complaint if the characters were supposed to be good listeners, say, friends trying to be supportive of the player’s own problems.

  16. Matthew Johnson Says:

    The Eliza ref in Terminator was in one of the last episodes — the second last, maybe? — where she and John have a conversation in his room (sorry I can’t be more specific). Basically, she talks to him as though she were an Eliza program — inverting things he says to her, and saying things like “And how do you feel about that?” It was subtle, but there was one key phrase that was such an Eliza mainstay it clinched it for me. (Alas, I’ve forgotten what the phrase was!)

  17. scott Says:

    Those of us who have experienced the “hot coffee” sequence in Facade know that it already kicks the shit out of ELIZA. Thanks for that easter egg. It keeps me coming back for more.

  18. andrew Says:

    Ah, Scott, that explains why you were “banging air” in the AR Facade exhibit at the Beall show.

  19. scott Says:

    Teledildonics rock!

  20. Mark M. Says:

    So the question remains:

    Are these features of context part of the data or part of the processes? I think these comments prove that you can’t have one at the expense of the other. In my book project, I develop a reading of Michael Mateas’ addressing the context as the realm of formal affordances and the code logic as the realm of material affordances. The latter is the set of possible input. The former is the set of input that will make sense given the developing scenario. (You can see some of this in my dissertation).

    Again, for me the literary nature of this context and data is one of the barriers to entry into my category of electronic literature. Once that criteria is met (and in light of that context), I can consider the roll of the processes. But again, I feel like my position is just an inverted set of values of the one you presented in your paper.

  21. nick Says:

    Mark, this seems like an important issue, but I’m not sure what “features of context” you’re talking about. Terminator references? Help me out here?

  22. Mark M. Says:

    The features of context would be any aspect of the bot that signal establish its context. So we’ve been talking a lot about the ways in which bots situate themselves (and their failings). ELIZA as Doctor offers a conversational context of an exchange with a Rogerian therapist. The Prayer bot presents the exchange as prayers offered to a deity. Facade, and my chapter goes into more detail here, lays out a very particular constraint of an encounter with two friends from college at the apartment 10 years after college within the confines of a 15-20 minute 2 act tragedy. As you can see with my description — that context is signaled by data, logic, and presentation layers together.

    My guess (and my experience from trying to create bots) is that the conceit of this narrative context — scripted into all the data and built into the presentation layer (to give meaning to the logic or processes of the bot) is what’s key to making a satisfying bot. Satisfying to whom? the reception theorist might ask. Satisfying as literature? we might ask. But for now, I’ll just say satisfying to me.

  23. nick Says:

    I see – you mean that Eliza enacts/impersonates/plays/parodies a Rogerian psychotherapist. I think I would probably call that Eliza’s role or script rather than its context. That seems to me to be part of the “text” of Eliza/Doctor, broadly speaking, not what surrounds and situates the text. On a higher level, Eliza is an interactive computer program rather than, for instance, a character in a play who is a psychotherapist. But I think even at this point we are talking about the form and not about the context.

    The context seems to me to be the overall cultural and individual situation in which a person has some idea of what a psychotherapist is. This could come from being in therapy, from popular representations of therapy, or perhaps from being a therapist. And, at the level of form rather than a particular role, there is a context as well. We also have expectations about and understandings of what a computer program is and what a play is.

    As for the question, What makes Eliza impersonate a psychotherapist – data or process?, both are clearly important. By making point 6 about process intensity, we certainly didn’t mean to say that data is irrelevant, only that process is more relevant than has been acknowledged and more relevant than it is in most e-lit. Eliza/Doctor has a particular role because the program reflects the interactor’s language back (mainly due to process) but also because the program produces neutral, gentle language (mainly due to data). In contrast to most electronic literature, “content” is not king with Eliza. Modified versions of the program with different response texts, not preserving the original wording of Weizenbaum’s first program, are still “Elizas.” But that isn’t to say that data is irrelevant. If you were to change just the data and have Eliza hurl obscenities or obsessively mention cheese, that would scuttle the program as a psychotherapist impersonator and make it into something else.

  24. Ian Bogost Says:

    One important aspect of Eliza, which informs both data and process, is its material context. The teletype interface that Eliza uses already sets a whole range of expectations from the user, including the idea that he or she might be talking to a human agent via a timeshare computing system. I think Nick also once observed (somewhere, where NIck?) that the program’s slow output via teletype encouraged even more gap-filling than do the modern, immediate screen versions.

    What modulates the user’s expectations here is not just the design of the system, but the tight coupling between that design, its material context, and its social/cultural situation. Thus the (possibly apocryphal) story of the VP who mistakenly thought Eliza was a truculent employee, or that of Weizenbaum’s secretary, who eventually asked him to leave the room while she “spoke” to the program, despite knowing full well it was a program.

  25. nick Says:

    Ian. that’s in my paper “Continuous Paper” which is online in a few different versions … here’s the ISEA “Continuous Paper,” which Scott graciously presented for me.

    Interestingly, a slower pace of response, more imitative of human conversation than modern-day computer reply, can also be found in a system like the Apple II running some interactive fiction program that has to go to disk all the time. But it’s certainly a noticeable feature of systems with a teletype interface.

  26. Game AI Roundup Week #24 2008: 13 Stories, 4 Quotes, 1 Video, Job — Says:

    [...] Provocation by Program: Imagining a Next-Revolution Eliza [...]

  27. robert-depot » Blog Archive Says:

    [...]… [...]

Leave a Reply

Powered by WordPress