Whether CTPs should walk on three legs or two; how the robotic artwork Petit Mal is “interpretationally plastic;” what cultural assumptions we build into machines: just some of the response-topics here.
Phoebe Sengers responds in turn
Phoebe Sengers responds in turn
(To Lucy Suchman)
Yes, life is not a story, and therefore, by extension, the Patient is not alive. The lesson from the critique from anti-psychiatry and narrative psychology for Artificial Intelligence is the danger of assuming that life can be cleanly understood and implemented, represented, as Lucy Suchman says, by an “omniscient and regulatory `designer’-creator.” This is not to say we cannot tell stories of life, but that those stories must always be traceable to the one who is doing the telling, who is accountable for what those stories say. In the nightmare world of the Patient, therefore, all trappings of the omniscient nature documentary have been removed; the Patient is a cartoon, the Patient is fantastic, the Patient was clearly built by someone with a goal in its telling, and the Industrial Graveyard and its user-unfriendly interface are designed to make this clear. The Patient is, in the tradition of literature and film, enacting a story; it is, in fact, my story [Sengers, 1995], which may throw a new light on the confusion that Suchman feels!
At the same time, one must ask whether the construction of the Patient fully allows for the flexibility, multiplicity, and negotiability of narrative. The answer to that is no. One of the major shortcomings of the system is that I serve up prepackaged narrative without much leeway for audience interpretation. I decide on all the behaviors and transitions ahead of time, and then the goal of the system is simply to make sure that those decisions make it across the yawning divide to the user intact.
A different approach that is friendlier to the value of negotiability is that taken by Simon Penny in his robotic artwork, Petit Mal [Penny 2000]. The design of Petit Mal explores the extent to which people can attribute meaningful behavior to autonomous robots. Petit Mal is set up, not to elicit any particular behavioral interpretation, but to allow for many possible behavioral interpretations. Far from trying to impose particular interpretations on the user, Penny uses Petit Mal as a blank screen onto which many possible interpretations can be projected. Petit Mal is interpretationally plastic, and never exhausted by the onlooker’s musings; this gives its dynamics a degree of liveliness that the Patient lacks.
The difficulty with this plasticity is that it is relatively low-level. At the internal level, Petit Mal does some simple navigation and obstacle avoidance (which is, of course, regularly interpreted as much more complex behavior). It is not clear how much more complex behaviors can be constructed for Petit Mal without simultaneously greatly constraining the interpretational space. In this sense, Petit Mal and the Patient occupy more or less opposite ends of the spectrum of interpretational negotiability on the one hand and understandable complexity on the other. If this is so, it might be interesting to now try working towards something in the middle.
Still, there is something deeper behind the strain which Suchman identifies: it is a strain between the disciplines, the question of whether intellectual labor in the field of cultural critique can be made meaningful within the tradition of agent-building practice and vice versa. Following Suchman’s admonition might mean to drive the project largely by the cultural critique involved, to take an outsider perspective on agent-building and to suggest that current agent-building traditions carry so much negative cultural baggage that the only hope is a radical new practice and a complete break with tradition. Although such critique can be useful, I believe it is essential, at least some of the time, to be willing to enter the looking-glass world of agent-building with one’s cultural armory in tow, and to be able to relate the insights of cultural analysis to “that world’s characters, problems, projects and prospects.” Without this willingness, one cannot expect those involved in agent-building traditions to listen to and learn from the critical perspective. The cost of this intellectual border-crossing is that there is not one neat, coherent perspective from which the world can be viewed. This costs the reader effort, as Suchman has noted. Integrating two such widely divergent worldviews and conversations is likely possible only on a contingent basis, driven by particular problems and projects. Nevertheless, I think it is an effort well worth taking.
(To Michael Mateas)
Philip Agre’s formulation of critical technical practices opens a rich space of possibility by stating clearly that there is not a single form of critical technical practice, but multiple possible such practices. In this context, the questions Michael Mateas raises about the relationships among different critical technical practices - what properties they share, what they can learn from each other, in what ways they are, perhaps, incommensurable - are essential to ask in order to understand where this field is going. I thank him for the opportunity to address these issues. Here, I have organized my answers to his questions according to two rubrics: the nature of critical technical practices in general, and the function of socially situated AI as a specifically agent-oriented practice in particular.
The nature of critical technical practices
In order to be able to talk about the relationship between Socially Situated AI (SS-AI) and Expressive AI (E-AI), we need to understand how they each relate to Agre’s original notion of critical technical practices (CTP). In Agre’s formulation, in a CTP, problems are encountered at a technical level, then understood and addressed as philosophical problems, and finally resolved at a technical level on the basis of philosophical insights. “The point… is to expand technical practice in such a way that the relevance of philosophical critique becomes evident as a technical matter ” (Agre, p. xiii). Important to note here is the primacy of the technical over the critical in a critical technical practice. One must first start with a technical problem, then one can take a critical or philosophical approach, by which one finds a technical solution.
This is, in fact, true of SS-AI, at least the work on expressive agent architecture which I describe in First Person. It is not, however, true of Expressive AI. In Expressive AI, the opposite situation holds: the technical problems that the artist chooses to tackle are a consequence of the artist’s vision of what it is he or she would like to communicate. It is, of course, tempting at this point to crow about the superior authenticity of SS-AI in this respect, but, in fact, Mateas’s work is a good example of why, with apologies to Agre, I believe it is unfruitful to take the primacy of technical work as essential for CTPs as currently practiced.
Although Agre’s work in the area derived from work in AI practice, and he sees critical technical practices as first and foremost a variation of a technical practice, what seems constitutive about it is not that the problems come from technical work originally, but that work proceeds simultaneously or alternately from a technical and a philosophical perspective. If, indeed, in a CTP technical and philosophical problems are seen as two sides of the same coin, then there is no a priori reason why the technical problem should come first; one can instead start with a philosophical problem, which then becomes instantiated as a technical problem. Wardrip-Fruin and Moss make a similar point. And with nominally digital art practices (such as those of Simon Penny and Chris Csikszentmihalyi) and nominally design practices (such as those of Bill Gaver and Tony Dunne) approaching asymptotically the work of more officially technical practices (such as those of Mateas and myself), it seems counterproductive to stick with the primacy of technical work as a defining characteristic. The important point from my perspective is that the CTP make coherent sense as one practice both from a technical and from a philosophical or critical perspective.
In addition to this requirement, Mateas argues that there needs to be at least three disciplines involved in a CTP. I am not sure if I disagree about the principle or about the definition of what it takes to be a discipline. My own practice as described here may stand on the three legs of AI, cultural theory, and narrative psychology, but if that is the case it is a very wobbly practice, as the leg of narrative psychology is much shorter than the other two! I see narrative psychology as simply one part of my cultural theory practice; one can see this truly as three legs only if the critical perspective is separated from the rest of the cultural studies content, which seems artificial, but might be useful for thinking about how CTPs function in general. Instead, I propose the following formulation: that a CTP can be the synthesis of a technical and any other form of practice, where the second practice, like art or cultural studies, has a critical perspective as an essential component, and where that critical perspective is brought to bear on the technical work (this is not intended to rule out three-pronged structures, or a minimalist CTP with only a technical and `pure’ critical perspective at work, though it does seem likely that critiquing technical approaches without having another discipline to draw on for a concrete alternative would be difficult).
The primacy of agency
As Mateas points out, SS-AI is limited in its definition to agents - perhaps a better name would be “Socially Situated Agent Design.” “Agent” is not intended here as a catch-all term for all of AI. I do not see the focus on agents as an inherent limitation; rather, I see SS-AI as a case study in a particular domain, from which analogies could be drawn to other areas of AI. At the same time, agency is a particularly important case study for the principles of SS-AI, because it is arguably the part of AI where the postulates are most likely thought not to hold by practitioners. Agents, unlike other AI and non-AI systems, are theorized as fully autonomous, i.e. as existing independently of their creators and audience. As such, agents form an extreme case; it follows that other AI systems might be more easily socially contextualized.
Yes, the postulates of SS-AI hold if we replace “agent” with “AI System.” In fact, the postulates still hold if we replace “agent” with “computational system” - in which case we end up with the postulates of Human-Computer Interaction. In this sense, SS-AI is simply stating that what HCI practitioners have known all along should also be the perspective of agent builders.
SS-AI is one example of a CTP, not a definition of what a CTP in AI must look like. In my more recent work, I have been adapting my experience with SS-AI to other domains. My work in the area of avatars, or agents which are intended to represent human users, has similarly been based on recontextualizing the nature of autonomy in the human-avatar relationship [Penny, et al.]. My current work in the area of Ecological Media, or smart appliances which support awareness of the environment in day-to-day activity, is not focused on agents at all, but rather on changes in perception that interactive media can bring about. In working in these areas, the following postulates generalized from SS-AI have proven repeatedly useful:
1. Systems need to be evaluated not only within the technical frame within which they have been defined, but also with respect to the social and cultural environment that shapes the system and which the system affects.
2. Systems encode culturally specific assumptions. These assumptions have material effects in the behavior of the system and are experienced by the user of the system.
3. Because of this, responsible system design involves careful selection of the cultural assumptions built into the machine.
In response to Suchman
Penny, Simon. Agents as Artworks and Agent Design as Artistic Practice. In Human Cognition and Social Agent Technology, ed. Kerstin Dautenhahn. Amsterdam: John Benjamins Publishing Company, 2000.
Sengers, Phoebe. “Madness and Automation: On Institutionalization.” Postmodern Culture. 5(3), May, 1995.
In response to Mateas
Agre, Philip E. Computation and Human Experience. Cambridge, UK: Cambridge UP, 1997.
Penny, Simon, Jeffrey Smith, Phoebe Sengers, Andre Bernhardt, and Jamieson Schulte. “Traces: Embodied Immersive Interaction with Semi-Autonomous Avatars.” Convergence. Vol. 7, No. 2, 2001.
Wardrip-Fruin, Noah and Brion Moss. “The Impermanence Agent: Project and Context.” Cybertext Yearbook 2001, ed. Markku Eskelinen and Raine Koskimaa. Saarijärvi: Publications of the Research Centre for Contemporary Culture, University of Jyväskylä, to appear.