Wednesday, September 05, 2018

Reading :: Human-Machine Reconfigurations

Human-Machine Reconfigurations: Plans and Situated Actions, 2nd Edition
By Lucy A. Suchman


Lucy Suchman's Plans and Situated Actions (1987) was a huge influence on the fields of human-computer interaction and computer-supported cooperative work. It currently has 11586 citations, according to Google Scholar, and for good reason: This book changed HCI by demonstrating that the then-dominant view—the view embraced by information-processing cognitive psychology, in which an abstract mind engaged in cognition-as-computation—could not adequately account for how people engaged with machines. That is, when working with machines, humans did not use plans (in the sense of stepwise series of abstract actions) so much as situated actions (in the sense of local, sensed alternatives at each moment). Based on her ethnomethodological studies at Xerox PARC, conducted for her PhD in anthropology, Suchman demonstrated that photocopiers' help systems were built for plans, but their users worked through situated actions, resulting in mismatches and rendering the help systems unhelpful.

I discovered this book early in my PhD program (1994 or 1995) and spent a while with it and similar books, as well as the PDFs of technical reports supplied by Xerox PARC and EuroPARC. (The WWW was relatively new then, and the idea of downloading research papers directly from the research institution seemed magical.) Suchman, who was a PhD student at the time, wasn't alone—others with interests in ethnography, ethnomethodology, and action research also worked or interned at these institutions, including ethnomethodologist Graham Button, linguist Geoffrey Nunberg, and computer scientist Susanne Bodker, whose 1991 Through the Interface (a foundational text for activity theory in HCI) was based on her 1987 dissertation. In fact, the late 1980s and early 1990s yielded many texts that questioned the then-dominant information-processing cognitive psychology in HCI and introduced constructivist perspectives. But Suchman, deservedly, became perhaps the best known.

What's the relation of that 1987 book to this 2007 book, Human-Machine Reconfigurations? It's complicated. On one hand, Plans and Situated Actions is in here as Ch.2-10, nearly untouched (some of the footnotes have been updated). On the other hand, the new book has a new Ch.1, "Readings and Responses," and a new Ch.11-15. That's an additional six chapters, more than one would expect from a second edition, but less than one might expect from a new book. The new title and subtitle reflect the mixed message.

To be honest, the core of the original book holds up, and Chapter 1 works to situate it as a historical document, but I did not think that the new chapters at the end were of the same quality. Since I haven't reviewed the original Plans and Situated Actions on this blog, I'll spend most of my time in this review on it.

In Ch.1 (a new chapter), Suchman gives us the background of the original study from her viewpoint 20 years later. That study began "in 1979, when I arrived at PARC as a doctoral student interested in critical anthropology of contemporary American institutions and with a background as well in ethnomethodology and interaction analysis" (p.8). She became interested in interactivity when her colleagues attempted to design the interface for a new photocopier. The copier had been advertised as so simple that one had only to press a green button (p.8)—but customers complained that it was too complicated (p.9). To investigate, she videotaped her colleagues attempting to work with the machine, and on that basis, concluded that "the machine's complexity was tied less to its esoteric technical characteristics than to mundane difficulties of interpretation characteristic of any unfamiliar artifact" (p.9). Based on this study, and on her understanding via ethnomethodology that contra AI assumptions "human conversation does not follow the kind of message-passing or exchange model that formal, mathematical theories of communication posit" (p.10), she essentially applied conversation analysis to "people's encounters with the machine" (p.10). She noted that when she was in the room, she could see how she "might have intervened," but "the machine appeared quite oblivious"—so "what resources was I ... a full-fledged intelligent observer, making use of in my analyses" compared to those of the machine? "The machine had access only to a very small subset of the observable actions of its users" (p.11). "My analysis, in sum, located the problem of human-machine communication in continued and deep asymmetries between person and machine" (p.11).

This problem is more complicated than her colleagues assumed, she says, because "I take the boundaries between persons and machines to be discursively and materially enacted rather than naturally effected and to be available ... for refiguring" (p.12).

Suchman then gets to her distinction between plans and situated actions. "My position then and now has been that plans are conceptual and rhetorical devices (often materialized in various ways, as texts, diagrams and the like) that are deeply consequential for the lived activities of those of us who organize our actions in their terms" (p.20), but they do not constitute a stepwise program. Here, she addresses some of the pushback that she received after the book's initial publication, regretting a word choice that caused some readers to misinterpret her argument.

From here, we get into the original book. In Ch.4, Suchman overviews interactive artifacts as they have been historically treated in computer science, noting that in the early 20th century, in the name of turning the study of cognition into a science, "the study of cognition as something apart from overt behavior was effectively abandoned" (p.36); the strategy in cognitive science in 1987, and especially in AI research, was to reduce cognition to computation (p.37).

In Ch.5, "Plans," Suchman discusses AI's consequent understanding of plans: "plans are prerequisite to and prescribe action, at every level of detail," and "mutual intelligibility is a matter of the reciprocal recognizability of our plans, enabled by common conventions for the expression of intent and shared knowledge about typical situations and appropriate actions" (p.51). That is, they are programs: "The planning model in cognitive science treats a plan as a sequence of actions designed to accomplish some preconceived end," where "action is a form of problem-solving" (p.52) and "actions are described, at whatever level of detail, by their preconditions and their consequences" (p.53). "Goals define the actor's relationship to the situation" and "the plan is prerequisite to the action" (p.53). Notice that this understanding of planning assumes an individual actor. In interaction, the model is extended to 2+ actors, and others' actions must be seen as expressions of their underlying plans (p.56). As Suchman notes, the literature equivocates "between plans as a conceptual framework for analysis and simulation of action and plans as a psychological mechanism for its actual production" (p.58). Yet the relationship between intent and the actual course of action is "enormously contingent" (p.60). This situation is not helped by AI's uptake of speech acts, which takes Austin's claim that "language is a form of action" to mean that communication can be subsumed to the planning model (p.61). Here, the problem of inferences is handled through "scripts" (p.64).

In contrast, Suchman embraces Garfinkel's view: "a background assumption ... is generated by the activity of accounting for an action when the premise of the action is called into question" (i.e., post hoc) (p.67). Thus "plans and goals do not provide the solution" for the problem of interaction; "they just restate it" (p.67).

In Ch.6, Suchman overviews the contrasting notion of situated actions, drawn from anthropology and sociology (p.69). In this view, plans are "resources for people's practical deliberations about action" that are "located in the larger context of some ongoing practical activity" (p.69). (In the late 1980s and early 1990s, it was popular to characterize the larger context as a way to get past cognitive science's stepwise focus.) Suchman asserts that "every course of action depends ... on its material and social circumstances" and her aim is "to investigate how people produce and find evidence for plans in the course of situated actions" (p.70). To do so, she turns to ethnomethodology, which she overviews in the remainder of the chapter.

Let's skip a bit here, since the actual study data are not that interesting for us here in 2018. Suffice it to say that Suchman demonstrates that the cognitive science understanding of plans is not adequate for explaining the frustrating encounters with the copy machine that her users encountered, and that those of us who have used copy machines recognize.

Now we get into the new areas of the book, Ch.11-15.

In Ch.11, "Plans, Scripts, and Other Ordering Devices," Suchman surveys literature on ordering devices that has come out since the original book. This literature survey includes John Law, Liam Bannon, Phil Agre, Steve Woolgar, Madeline Akrich, and others who will be familiar to those who read this blog.

In Ch.12, "Agencies at the Interface," Suchman surveys literature on the so-called "smart machine," starting with ALICE and ELIZA.

In Ch.13, "Figuring the Human in AI and Robotics," Suchman surveys the literature on humanlike machines.

Chapter 14, "Demystifications and Reenchantments of the Humanlike Machine," surveys encounters with humanlike machines.

Finally, in Chapter 15, "Reconfigurations," Suchman says "In this chapter I consider some new resources for thinking about, and acting within, the interface of persons and things" (p.259).

As you can tell from my characterizations of the new chapters, I didn't find much to write about in these chapters. I'm not sure why the author and publisher decided to release the second edition in this form, but the contrast between the two parts of the book is profound. In Plans and Situated Actions, Suchman is white-hot, incisively identifying fundamental problems in AI and HCI and deftly illustrating them with data that we can all recognize from our own interactions. In contrast, the expanded chapters are bland literature reviews that do not clearly relate back to the original argument, do not seem to advance the ball, and do not appear to make a contribution other than the survey.  In terms of argument, they don't seem to connect. In terms of genre, the literature reviews come at the wrong end of the book. And in terms of story, Suchman goes from being an MVP to a spectator.

None of this takes away from Plans and Situated Actions, which remains vital reading and is perhaps enhanced by Suchman's new context-setting chapter (Ch.1). For that reason, I still highly recommend the book (or its first edition, which is still on my shelf). But I wish that Suchman and the publisher had kept the original title and replaced the last five chapters with a brief afterword.

No comments: