12
250 BOOK REVIEWS 8. Polanyi, M., The Tacit Dimension (Anchor Books, Garden City, NY, 1967). 9. Pribram, K.H., Languages of the Brain (Wadsworth, Monterey, CA, 1977). 10. Rorty, R., Philosophy and the Mirror of Nature (Princeton University Press, Princeton, N J, 1979). 11. Schank, R.C., Failure-driven memory, Cognition and Brain Theory 4 (1) (1981) 41-60. 12. Weizenbaum, J., Computer Power and Human Reason: From Judgment to Calculation (Freeman, San Francisco, CA, 1976). 13. Wilber, K.W. (Ed.), The Holographic Paradigm and Other Paradoxes (Shambhala, Boulder, CO, 1982). 14. Winograd, T., Understanding Natural Language (Academic Press, New York, 1972). 15. Wittgenstein L., Philosophical Investigations (Macmillan, New York, 1958). On Understanding Computers and Cognition: A New Foundation for Design A response to the reviews Terry Winograd Computer Science Department, Stanford University, Stanford, CA 94305, U. S, A. Fernando Flores Logonet, Inc., Berkeley, CA 94704, U.S.A. 1. A theory of language In Understanding Computers and Cognition, we presented a theory of lan- guage, on which we base our understanding of cognition and of computers. It includes some basic assertions about how language works: (1) Language does not convey information. It evokes an understanding, or "listening," which is an interaction between what was said and the pre- understanding already present in the listener. (2) An utterance produces different understanding for different listeners, since each person has a background of pre-understanding generated by a particular history. This background does not determine interpretation in a rigid way, but generates the domain of possibilities for how what is heard will be interpreted. (3) The background that is relevant to understanding grows out of concerns, practices, and breakdowns in those practices. People interpret language in a way that makes sense for what they do. (4) The background of concerns and practices is not purely individual, but is generated within a tradition. Each person is unique, sharing background to varying extents with other people. Some amount is universally human; more is shared with members of the same culture; more yet with those in the same line of work; and still more with partners in frequent conversation.

Wino Grad 1987

Embed Size (px)

DESCRIPTION

Wino Grad 1987

Citation preview

250 BOOK REVIEWS

8. Polanyi, M., The Tacit Dimension (Anchor Books, Garden City, NY, 1967). 9. Pribram, K.H., Languages of the Brain (Wadsworth, Monterey, CA, 1977).

10. Rorty, R., Philosophy and the Mirror of Nature (Princeton University Press, Princeton, N J, 1979).

11. Schank, R.C., Failure-driven memory, Cognition and Brain Theory 4 (1) (1981) 41-60. 12. Weizenbaum, J., Computer Power and Human Reason: From Judgment to Calculation

(Freeman, San Francisco, CA, 1976). 13. Wilber, K.W. (Ed.), The Holographic Paradigm and Other Paradoxes (Shambhala, Boulder,

CO, 1982). 14. Winograd, T., Understanding Natural Language (Academic Press, New York, 1972). 15. Wittgenstein L., Philosophical Investigations (Macmillan, New York, 1958).

On Understanding Computers and Cognition: A New Foundation for Design

A response to the reviews

Terry Winograd Computer Science Department, Stanford University, Stanford, CA 94305, U. S, A .

Fernando Flores Logonet, Inc., Berkeley, CA 94704, U .S .A .

1. A theory of language

In Understanding Computers and Cognition, we presented a theory of lan- guage, on which we base our understanding of cognition and of computers. It includes some basic assertions about how language works:

(1) Language does not convey information. It evokes an understanding, or "listening," which is an interaction between what was said and the pre- understanding already present in the listener.

(2) An utterance produces different understanding for different listeners, since each person has a background of pre-understanding generated by a particular history. This background does not determine interpretation in a rigid way, but generates the domain of possibilities for how what is heard will be interpreted.

(3) The background that is relevant to understanding grows out of concerns, practices, and breakdowns in those practices. People interpret language in a way that makes sense for what they do.

(4) The background of concerns and practices is not purely individual, but is generated within a tradition. Each person is unique, sharing background to varying extents with other people. Some amount is universally human; more is shared with members of the same culture; more yet with those in the same line of work; and still more with partners in frequent conversation.

BOOK REVIEWS 251

The reviews collected here offer a striking validation of this theory. The same piece of language (in this case, our book) produced four widely different understandings, each generated within the background of a particular listener. The important issue is not that different reviewers " l iked" the book more or less, but that they heard it as addresing different concerns in different ongoing conversations. We will begin our response by examining their interpretations and the backgrounds in which they arose.

2. Four traditions

In the book, we use "tradi t ion" in a relatively broad sense, discussing at some length the "rationalistic tradit ion," which has developed over the course of centuries in Western society (see Chapter 2). In this response, we will focus on more narrowly delineated traditions, associated with a particular academic discipline or even a "school" within a discipline. At this scale the differences of background among the reviewers become most visible.

In identifying each of the reviews with a tradition, we do not imply that an individual can be taken purely as a representative. Each of us is a product of a unique history and has individual concerns. However, all too often in our culture, people focus on individuality without recognizing the degree to which thought and language are shaped by a history shared with others. In charac- terizing these four responses as representing traditions, we deliberately de- emphasize the personal aspects, in order to bring out the general i ty-- to show how these authors reflect a much wider circle of potential readers.

Vellino

Vellino responds in the tradition of academic analytic philosophy: his primary concern is the articulation of " ideas" and logical arguments supporting them. The abstractions take on a life of their own, and the discourse centers on their logic, rather than on trying to make sense of them with respect to some world of practical concerns. In his entire review, he does not consider any of the questions we take as primary: how our theory makes sense in the practice of work and what it offers for design. Instead, he enters into a debate about the arguments, focussing on identifying " . . . mistaken belief that the language of the analytic tradition is not sufficient to express ideas that are critical of its own presuppos i t ions . . . [and that] analytic theories cannot handle context depen- dency in natural language."

What is most revealing in these comments is the unquestioned assumption that the appropriate domain for discussion is that of "beliefs," " ideas," "presupposit ions," "analysis," " theory , " and so on. One of the major direc- tions of our book is precisely to question this assumption-- to show how language works in terms that do not postulate such objects of thought (see Chapter 5). Of course, Vellino is not alone in using this language and the form of argumentation that goes with it. It is a central thread in the rationalistic

252 BOOK REVIEWS

tradition, in which all of us are immersed. The style reflected in his comments is so pervasive in academic discourse that it is hard to recognize it for what it is: an extremely useful, but also limited way of understanding.

We see the same concern in Stefik and Bobrow's review. For example, they assert that "Projections of the ultimate scope of a technology should be argued from carefully defined assumptions and sources of fundamental limits," adopt- ing the popular model of the physical sciences and assuming that it is the only appropriate mode of discourse. They adopt the same narrow focus on defini- tion in trying to understand the rationalistic tradition:

Search for a definition leads to puzzlement about the rationalistic tradition: It is a strange composite that includes one rendition of the scientific method, some out-of-favor notions about how sent- ences convey meaning, the physical symbol system hypothesis in AI, and various kinds of reductionism . . . . It is not clear to whom they ascribe this mess.

There is a basic failure here to recognize how traditions are manifested. They are not coherent bodies of carefully articulated theory, of the kind that might be ascribed to individual authors and defended in a debate. In talking of a tradition, we are not identifying and labelling a th ing--a "school" or "s tance" or "posi t ion." We speak of tradition to point out how our history of conversations (both individual and in a society) shapes our language and thought, by providing the ground on which we work. As an analogy, consider the tradition of " f ree-market economics" that is prevalent in present-day America. One might say:

Search for a definition leads to puzzlement about f ree-market economics: It is a strange composite that includes one rendition of the functioning of pure competit ion, some out-of-favor notions about how government monetary policy can guide the economy, the welfare ideas of the New Deal, and various kinds of ant icommun- ism and social Darwinism . . . . It is not clear to whom they ascribe this mess.

The answer is, of course "To nobody- - and to all of us." We participate in a tradition and it changes through our participation. But we do not choose it or design it. It would be foolish to ignore the power of this particular tradition (as distinct from a European Social-Democratic, or a Chinese Communist one, each of which would be equally complex) because it cannot be precisely defined.

Special effort is required to recognize one 's own tradition and to question what is taken for granted within it. We have a t tempted to do that with the rationalistic tradition. In its a t tempt to reduce everything to logical arguments among competing ideas and beliefs, there is a constant danger of ungrounded abstraction and a kind of "bl indness." Rather than listening to language for

BOOK REVIEWS 253

what it might mean or do, one jumps eagerly into the game of "making points," "defending positions," and searching for "mistaken beliefs." Stefik and Bobrow complain that we "make little attempt to present a balanced or representative account of the relevant ideas." Indeed, that is not what we were trying to do. The very idea of a "balanced or representative account" takes for granted a broad consensus as to what constitutes the domain of "re levant" discourse. Our book aims to move outside of the conventional discourse--to stimulate new thinking about computer design, not to set up and defend a position (or what Vellino calls a "conceptual scheme.")

Stefik and Bobrow

Stefik and Bobrow are researchers who create programming techniques for use in artificial intelligence. Their concern for the logic of arguments is in service of their concern with inventing new theories and mechanisms. They find it disturbing that we criticize current research without offering alternative ways to proceed in their enterprise. If we argue that current techniques cannot do what AI researchers have claimed they will, then what other techniques do we advocate? Instead of answering that directly, let us examine the research paradigm that they represent, which we characterize at length in our book (see Chapters 7-9) .

Put in a simple slogan, AI research is the quest to capture the essence of thought and language in the form of explicit symbolic representations. The statement by Stefik and Bobrow quoted above about "carefully defined assumptions" is indicative of the implicit assumption that everything worth saying or thinking can be so defined. Further, Stefik and Bobrow belong to a predominant tradition within AI that sees little value in foundational (called "philosophical") questions. Shortcomings of current techniques and systems are seen as a transient failure to cover the right details and to invent the right clever mechanisms, rather than as anything more basic. Instead of a radical challenge to foundations, they would rather see arguments over the technical details: "I t would have been more interesting to read a critique of Grosz's work on comprehension using the context of a discourse, or Cohen's . . . . It would have been more useful if they gave a more balanced and comprehensive account of current research directions."

They imply that if we focus on the details of the technology, somehow it will all work out. They indulge in science-fiction speculations, such as " . . . newer computer s y s t e m s . . , might acquire background that we humans bring to bear to understand the w o r l d . . . " and "Perhaps we shall one day develop classes of complexity for computational beings that relate in natural ways to Piaget's stages . . . . " and they evoke images of robots from the novels of Isaac Asimov. This kind of innocent optimism about technological achievement has been common among AI researchers since the beginning of the discipline. We hope that a major result of our book will be to call it into question and to open conversation about it. In doing this, we are not trying to halt or slow the

254 BOOK REVIEWS

development of computer technology. We accept technology as a basic fact of human activity, not inherently good or bad. We are concerned with its direction and its grounding, and we want to increase our collective capacity of thinking about it. Stefik and Bobrow state th~,t in our book they " . . . felt an anti-technological bias." Clancey was listening to us more clearly when he said "This book is anti-illusion, not anti- technology." It is dangerous to fall into naive optimism, believing that local successes at satisfying particular tech- nological objectives (such as building a micro-chip or a successful chess playing program) indicate a more global success at understanding what we are doing and where it is leading. In the face of technological illusions with negative practical consequences, it is necessary to be anti-illusion (and anti-delusion).

Suchman

Suchman [3] has examined practical situations of computer use, applying her training in the methods of the social sciences, particularly anthropology. Within that discipline there has been an ongoing foundational debate about what it means for a person from one tradition to understand or describe the practices and languages of a different one. The Western anthropologist enter- ing a remote native village presents an extreme case of different backgrounds leading to different understandings and interpretations. As Suchman points out, the "ethnomethodologis ts" recognize that the same problem arises in every at tempt to articulate a description of human culture and behavior. We need to become aware of what we take for granted in our own tradition, in order not to project it blindly onto what we observe.

The conversation called "e thnomethodology" is akin to the one that has gone on under the label "hermeneut ics ," and is much closer to our own theories of language than to those of the rationalistic tradition. It is no surprise that Suchman is more open to our critique and more able to recognize the ways we use language. After noting, as Stefik and Bobrow did, that there is no simple definition of the rationalistic tradition, she says (with apparent approv- al) "The only excuse for such gross generalization is the possibility that what is common is more important , for certain purposes than what is distinct. The purposes in this case are to take off one set of eyeglasses and try on another: to imagine what other way we might l o o k . . , at the phenomena of cognition that we are all trying to understand."

She sees our work as providing an "alternative v i e w " - - a way to alter our stance as an observer, and thereby to escape from previous blindness and to situate our understanding of computer technology in a context of human practices. This is appropriate and it is also misleading. In a way, the ethno- methodologists are t rapped in the domain they criticize. Their critique is solid, but the response is to search for bet ter methodologies - - to create precise structures (or "views") which can provide a systematic method for dealing with background. Suchman says:

BOOK REVIEWS 255

We need to know all these things [what the setting is, how activities are organized, etc.] through systematic analyses. To get the latter, we need a sound methodology for the study of situated human practices, and the application of such study to the question of designing tools; a need that itself delineates a major interdiscipli- nary research programme . . . . [Winograd and Flores] offer neither a strong theory, nor a strong method with which to uncover or build [the sociological foundations of design.]

While we concur with her plea that questions of computer design not be approached from a purely technical orientation, we do not share the appeal to methodologies-- the creation of a systematic "study" that can be applied to questions of design. As we will discuss below, we are very concerned with developing practices for design, and in this we will use systematic methodologies (and even analytic logic). But taking the methodologies as an end in themselves is ultimately limiting in the same sense as the analytic tendency to take the arguments as an end in themselves. We want to expand our ability as observers, within a context in which we are not detached but are engaged in the practices we ourselves observe.

The difference between detached methodology and concerned involvement is apparent in Suchman's defense of current AI research:

While research in AI may not get us to an artificial intelligence soon, or even ever, in principle it can still contribute to our understanding of human intelligence, through its efforts to get closer. Pushing an approach to the limit, like the application of logic to the control of situated action, or the stipulation of common- sense knowledge, can clarify what remains outside of the ap- proach's grasp.

This statement is carefully hedged ("in principle"), and as an abstract principle of scientific methodology, it sounds fine. But what is it saying in this context? It appears as justification for the continuation of AI research along its present lines, even if the stated goal is not to build intelligent robots (as implicit for Stefik and Bobrow), but to "contr ibute to our understanding." We reject the implicit argument. If the basic concern is the design of technologies that are appropriate to human settings, then we need to look for possibilities that go outside the old directions, not spend our efforts on more sophisticated "studies" within the existing framework.

Clancey

Clancey is clearly the reviewer most affected by the book. He tells of transformations in his own thought and language, as he allowed himself to

256 BOOK REVIEWS

enter seriously into our discourse. It seems strange in light of all the negative characterizations in the first two reviews that he describes the book as "intelligent, measured, and instructive," and urges others to read it, as " . . . it might change how you think about the world." It is too easy to dismiss this in terms of personal preference or to make psychological speculations. The relevant question isn't whether he likes the book, but what background he brought to it, and how he listened within that background.

Clancey has been primarily concerned with the practical design of expert systems and of educational techniques associated with them [1]. He has worked with "experts" in various subject domains, trying to apply the methods of artificial intelligence to codifying their work. Our theory fits his experience:

. . . this [our theory] explains what is so patently obvious when you work with experts, namely that they have so much difficulty laying out consistent networks and describing relations among concepts in a principled w a y . . .

In reading the book he was looking for new possibilities, not for arguments: " . . . This analysis might lead to an entirely different teaching method . . . . " When he encountered our unfamiliar use of language (what Vellino called the "obtuse use of English"), he saw it not as a source of confusion, but as a potential for new interpretations:

While a cursory scan shows the book to be full of j a rgon- - thrownness, readiness-to-hand, shared background, blindness, breakdown, commitment-- these words turn out to be useful for retaining the message. Like Freud's jargon (e.g. ego, subconscious) these terms introduce a new language for thinking about familiar things . . . . This gives meaning to "things only exist through lan- gauge." Do thrownness, readiness-to-hand, shared background, blindness, breakdown, and commitment exist? For that matter, do the ego and the subconscious?

Although phrases such as "retaining the message" show his grounding in the old tradition, his use of new language throughout the review demonstrates the degree to which it has taken root in his own understanding. The problems he has encountered in his own attempts to apply AI to real situations are the ones we address in our book: What are we doing with computers? How do we teach people? How are we affected by technology? Clancey demonstrates from personal experience that our discourse is relevant to these questions.

3. Listening to the critiques

What about our own background for listening? How do we hear the language of the reviews for our own concerns? As suggested in various sections above,

BOOK REVIEWS 257

we did not set out to create and defend a position. We do not pretend to have produced a definitive treatise in which one can find the "correct" synthesis. Our concern is to participate in the generation of a theoretical basis that works for design. As the "computer revolution" unfolds, new domains of human action are being created and old practices are being changed. We share with most of the reviewers a basic concern with developing "cognitive tech- no logy"- -a computer technology that is grounded in an understanding of what people do and how they think. But we do not share the assumption that the background for this development is adequately grounded in the existing academic disciplines of philosophy, linguistics and artificial intelligence. We see the need for a radical shift, not just better accounts within the prevalent tradition. At the same time, answers will not be provided by jumping to hermeneutics or speech act theory as a replacement. It is just as limiting to take these as the "right conceptual scheme." Our goal is to evoke openness to a new discourse that will guide the technology of the future.

Much of what we say is already implicit in other parts of the larger tradition of our society, and for that matter in some of the computer system design that is already done. Devices such as automatic teller machines grew out of an implicit understanding of tools for conversation, graphic user interfaces provide "readiness-to-hand," and elements of this understanding are visible even in programs such as operating systems. We want to design computer technology in a more principled way. Before the theory of thermodynamics, people built steam engines with a fair degree of success. Good designers did it well, without explicit articulation of a theory. But with explicit theory we can do more than improve the old designs--we can create new possibilities.

With this as our concern, we need to recognize what happens when our writing is interpreted. We will look at three ways in which these reviews indicate something that is potentially wrong or missing.

Failure to evoke an openness to the discourse

The first condition for language to work is that the hearer be open to enter into a conversation with the speaker. No amount of logical or beautiful or stirring language makes sense if it is not listened to. There is no use talking if the backgrounds of speaker and hearer are too different to create open and serious listening---even if the words are read, they will be too distantly reinterpreted.

Clancey says, "Ironically, the book's new view of cognition is a little scary, making reasoning seem limited and out of our personal control ." He points out that many potential readers will react by not giving the book an open and serious reading. It is clear from the reviews that his concern is valid. For Vellino, at least, it is apparent that we did not evoke a commitment to go beyond making arguments in order to understand what we were about. He says of one section, for example, "Frankly, I do not understand what this passage says that cannot be stated more simply by the following sentence: 'meaning

258 BOOK REVIEWS

depends on a set of background beliefs, desires, and dispositions. '" We could enter into a discourse on why "set of background be l i e f s . . . " is a totally distorting way of characterizing the background we discuss. In fact, we do so any number of times in the book. But Vellino's pre-understanding is not sufficiently close to take that discourse seriously. Serious listening means something different from endorsing or criticizing the logic:

. . . look first for the apparent absurdities in the text and ask yourself how a sensible person could have written them. When you find an answer, I continue, when those passages make sense, then you may find that more central passages, ones you previously thought you understood, have changed their meaning. [2, p. xii]

Stefik and Bobrow are more confusing. On the one hand, their explicit characterizations indicate a reading that rejects what we say:

[The book is] disappointing both for those who expect a carefully reasoned position, and those who seek an articulate account of the way things should be . . . . What they have to say about limits . . . has little enduring value . . . . [They] fail to characterize the ultimate shortcomings of technology in terms of fundamentals.

At the same time, as Editor-in-Chief and Book Review Editor of the most important theoretical journal in artificial intelligence, they have given the book the unprecedented prominence of this set of reviews. Their actions reveal another kind of listening, in which they too are affected in deep ways by what we say, but cannot deal with it in terms of the rational arguments and critical judgments they feel compelled to offer. As researchers in artificial intelligence techniques, they sense the frustration and dissatisfaction that have beset the field. Having seen the limitations of what they call "the current technology" and opened themselves to deeper questions, they cannot return to the innocent optimism of the early days. But at the same time, from their positions as senior representatives for the field, they are called upon to defend the received position and to pass judgment on those who cast doubts.

One senses a mood of anger in their repeated criticisms that we have not made things sufficiently "clear" or given sufficient arguments. They experience the disorientation that comes from questioning the grounds on which current discourse stands, and there is a natural reaction of disagreement and defense. Their characterizations of our writing (such as "confusing fragments of argu- ment and rhetoric for substance") show that we did not succeed with them in establishing the groundwork of trust that turns disorientation into an openness to new possibilities. But they have reached the first stage of what Clancey describes in the process of reading the book. Fundamental concerns are touched, and although their assessments are primarily negative, the sense of confusion (and frustration) shows that the book has evoked serious concern.

BOOK REVIEWS 259

Suchman and Clancey were more evidently open in their listening. This is not to say we agree with everything they write, or that they have an openness that transcends their own background. They like all of us, are always moving in the tumultuous space between the traditions they embody and a new under- standing that is emerging. Their response is neither an argument nor a judgment , but a participation in a conversation. Overall , the people that will do something about our central concerns are people, like them, who work with designing computer systems for real settings.

Generation of wrong interpretation

Even for a serious and open reader, a text can be written in a way that generates a wrong interpretation. In saying that an interpretation is "wrong" we are not appealing to some objective standard of truth, but recognizing that the interpretation would lead to different grounding and consequences for action than we anticipate. There are clearly some parts of our book that have not been uniformly effective in creating an understanding close to ours.

One example we discussed earlier was the misunderstanding of " t radi t ion" that led Stefik and Bobrow to treat it as something that can be defined and justified. A more troubling example (because it was misunderstood by several of the reviewers) is our discussion of the Coordinator s , and the t reatment of commitment within it as a systematic domain.

The design [of the Coordinator] is surprising in several ways. First, it uses vocabulary associated with rich and complex human interac- tions (e.g., promises) with very specific technical m e a n i n g . . . [this] seems to violate the spirit of their use of vocabulary . . . . They ignore the lesson they preach in the rest of the book. Terms like promise, commitment and responsibility are mutually defined and used by people in different ways depending on circumstances.

Suchman also interprets what we say as contradictory:

The concern with making things explicit, moreover , seems in some ways to contradict Winograd and Flores' commitment to under- standing things in terms of their relation to an unarticulated back- ground of community and practice.

What has been missed is the difference between an observational account of a domain of practices, and a space of distinctions that generates the domain. Indeed, as Stefik and Bobrow assert, " . . . commitment and responsibility are complex notions with social and legal dimensions." One cannot give precise definitions for them or reduce them to any mechanical algorithm or implemen- tation, and it was not our intention to do so. We made no claim of writing

1 The Coordinator is a trademark of Action Technologies, Inc.

260 BOOK REVIEWS

about the question of responsibility, and in writing about "commitment ," we are not trying to define how the word has been used, but to create a domain of possibilities in which it is a distinction (see Chapters 11 and 12).

As an analogy, consider commerce and finance. An arbitrary formalized construct like "I owe the bank $4,000" in no way captures the meaning of property, debt or value. But the systematic structure of money, expressed as standardized conversations about finances, gives us a mutually recognized basis on which to structure our transactions. The point of making things explicit in the Coordinator is to generate such a mutually recognized base. When we label a particular utterance as a "request ," we make no pretense of dealing with the full human context of requesting. We are providing the utterer and listener with a simple way to make an explicit declaration within a mutually understood context of practices. When someone sends a "request ," everyone involved can anticipate that it will be followed by certain potentials for further action and breakdown, just as we can recognize that a price tag of $1.59 signals a potential for certain business transactions.

In claiming that the distinctions embodied in the Coordinator are universal, we are not implying a uniformity of cultural interpretation. We are identifying a universal constitutive feature of human social activity. People ascribe to themselves and others the capacity to make commitments-- to use language in a way that allows others to anticipate their future actions. There will be very different cultural interpretations of who is able to make and recognize what kinds of commitments, and how to interpret the consequences of breakdowns when the promised action is not carried out. But there is always the basic relation, in which one person can take seriously another 's language about "I will do . . . . "

From the standpoint of design, the issue is how to provide tools for operating in this domain. As Clancey discusses in his Section 3.7, computers are especially suited to dealing with the formalized structure that provides a space in which to make explicit acts. The Coordinator is just one example. Au- tomatic Teller Machines, mundane as they may seem, are another application. They implement a network of possible conversational acts in which the customer and bank employees can participate, making particular recurring patterns explicit and "ready-to-hand" (see Chapter 12).

Missing possibilities for action

In looking for what is wrong with the book, we are at least as concerned with what is not there as with what is there. Clancey notes that the book is "of strong theoretical interest, but the practical implications, for example, what expert systems might ultimately be able to do, are not clear." Suchman urges the development of a "prototype process for design." They are not playing the role of critic, but are inviting us (and the readers) to go further. We are at an early stage of a new conversation, and are just beginning to reveal its

BOOK REVIEWS 261

consequences. The links have not yet been made that will carry our overall theoretical orientation into the specific practices of designer. That is the task for the coming years.

In particular, there is a body of practice based on theories of representation that pervades computer system building. We would be foolish to argue that it should be ignored or rejected, and as Clancey points out in his Section 4.4, there is much work to be done in reinterpreting and reintegrating it. As a next step in this process, we are now beginning a theoretical study of current practice in building "expert systems," and have begun to elaborate a theory of representation based on the emergence of distinctions in a linguistic domain, in response to recurrent patterns of breakdown. Much more theoretical work needs to be done here, in conjunction with the practical applications.

In a similar vein, we are developing a design methodology that will aid in creating systems that are consistent with our theories of human action and technology. A methodology is a kind of "coaching"- -not a formula for producing a result, but a set of practices that can lead to appropriate questioning and to appropriate change. We want to shift concern away from the computer-technology focus of the standard "system development" methods towards a concern with the design of work and language, using computer tools (see [4] for a first step in this research).

Finally, we are engaged in the direct development of the theory through practice. We are gaining a great deal of experience with The Coordinator as a commercial product with thousands of users, and are designing computer systems that go beyond both its simple facilities and the simplistic logic of "expert systems." As with all inventions, whether they are "right" is not a matter for abstract debate, but of seeing how they work and what new ways of working emerge in using them. This is true of the theory as well. Success is not the result of fending off competing arguments, but of generating a new ground on which argument is carried out. This kind of shift cannot be immediate-- i t will be measured in decades and proceed in unpredictable ways. The coming years of experience will be our teacher and our most exigent reviewer.

REFERENCES

1. Clancey, W.J., From Guidon to Neomycin and Heracles in twenty short lessons, A I Magazine 7 (3) (1986) 40-60.

2. Kuhn, T., The Essential Tension: Selected Studies in Scientific Tradition and Change (University of Chicago Press, Chicago, IL, 1977).

3. Suchman, L., Plans and Situated Actions: The Problem of Human-Machine Communication (Cambridge University Press, Cambridge, U.K., 1986).

4. Winograd, T., A language/action perspective on the design of cooperative work, in: Proceed- ings Conference on Computer-Supported Cooperative Work~ Austin, TX, 1986.