Kenneth D. Forbus and Paul J. Feltovich. (2001). Smart Machines in Education: The Coming Revolution in Educational Technology

	Kenneth D. Forbus and Paul J. Feltovich. (2001). Smart Machines in Education: The Coming Revolution in Educational Technology. Menlo Park, CA: AAAI Press/MIT Press. Pp. 483 $37.95       ISBN 0-262-56141-7 Reviewed by Bryan R. Warnick University of Illinois at Urbana-Champaign January 22, 2003 Smart Machines in Education: The Coming Revolution in Educational Technology is clearly a book for those curious about advances in cutting edge educational technology. The editors of the volume, Kenneth D. Forbus and Paul J. Feltovich, have brought together an illuminating collection of essays that show how research in cognitive science and artificial intelligence (AI) is shaping technological development in education. The essays reveal what the next generation of educational technology might look like, and help us grasp some of the reasons why the technology might look that way. It is also a book that can be read on different levels. Certain essays have something of a historical twist, tracing the negotiation that often occurs as educational technologies are developed, evaluated, and inserted into classroom environments. Other essays focus more on the theories of learning that undergird the technological applications; indeed, some of the essays do not mention technology at all, and instead focus on ideas about learning that might be more fully incorporated into future technological artifacts. Still other essays are of a more technical bent, relaying in some detail how their "smart machines" work. Thus, sections of this book should be of interest to a fairly broad segment of the educational research community. In what follows, I give a summary of the essays contained in this volume, and outline what I find new and intriguing about each technology. This is a lengthy book of almost 500 pages, so much more could be said about each article. After surveying the contents of the book, I will turn to some of the major themes that pervade the work, and discuss some of its underlying assumptions. Smart Machines: A Brief Summary The first chapter, "Representational and Advisory Guidance for Students Learning Scientific Inquiry," by Dan Suthers and others, reviews the authors' attempts to help students gain a feel for scientific argumentation through an application called BELVEDERE. In BELVEDERE, students are presented with a scientific problem (e.g., the extinction of dinosaurs) together with some information about research relating to the problem. The students then use this information to build a representation, or an "evidence map," of the scientific debate, by organizing the scientific statements and relations between the statements (e.g., showing which data support which hypothesis). A computer advisor helps students to reason through their evidence maps when help is needed. The authors relate in some detail how they developed this advisor and how through "evidence pattern strategies" and "expert-path advice strategies" the appropriate advice is offered. The second essay is entitled "Motivation and Failure in Educational Simulation Design" and is the work of Roger Schank and Adam Neaman. The authors describe simulations that present "Goal Based Scenarios" aimed at helping students gain expert skills in a "learning by doing" fashion. They first describe a simulation called Wolf, in which students play the part of wildlife biologists attempting to determine the cause of a declining wolf population. The idea is that students will learn science by actually going through the steps of scientific research. As the students develop their hypotheses, they can, at any time, select a question to ask a "wolf expert" and watch a video clip that shows an expert wildlife biologist offering a "war story" from his or her experience. Also included are exciting clips of a wolf being captured for examination and maps of prey populations. The most impressive thing about this application, however, and the others described by Schank and Neaman, is the care that the developers have taken to make user mistakes meaningful. Specifically, they try to simulate those conditions under which novices tend to make errors and then offer the appropriate just-in-time expert advice. Schank and Neaman include in their essay a thoughtful discussion of how to use failure to make errors both educative and motivational, and make a strong case that computer simulations can help students make educational mistakes in a non-threatening atmosphere. The next essay, "Technology Support for Complex Problem Solving: From SAD Environments to AI," by Guatam Biswas and others, discusses technology created to help improve students' problem solving abilities. Their approach to technological development is to begin with fairly simply technologies (SAD—stone age technologies) and add more sophistication as needed—a fruitful way, it seems, to avoid needless technological complexity. As an example of this developmental process, they show how they developed the original Jasper Woodbury problem solving adventure series, a fairly simple simulation in which students try to solve various problems. They found that students were having a difficult time transferring their newly acquired problem solving skills gained from this single, video-based activity to other problem domains. To solve this difficulty, they created Adventure Player that, with the help of more advanced technologies, attempts to create an environment that more fully facilitates the transfer of skills. They do this by providing additional features such as "what-if scenarios" that ask the students to use their new problem solving skills in new ways. Students can also see simulated results of their work, toil with a "coach" if they get stuck, or create simulated problems for other students to solve. Most provocative, though, was the second part of the essay that builds on the idea that we learn best when we teach. After some reflection on what, exactly, may be educative about teaching, the authors describe several ways learning is actualized through teaching, and then describe programs that allow students to teach a computer character some knowledge or skill. In one simulation called Betty's Brain, for example, students are asked to construct a brain for Betty that will then be able to answer questions about a river ecosystem. The students build a concept map with nodes (fish, bacteria, etc.) and relationships between the nodes (fish eats bacteria), and then test Betty's knowledge for accuracy. Betty can even be given different learning attitudes and attention levels to help students reflect on their own dispositions as learners. The idea is that if students can see as they teach what sort of learner characteristics are helpful, they can improve their own learning dispositions—it will be interesting to see how well this idea works in practice. The authors conclude their essay by pointing to the evidence showing that the learning-by-teaching strategy is having a positive effect on student learning. Chapter 4 is entitled "Growth and Maturity of Intelligent Tutoring Systems: A Status Report," and is the work of Beverly Park Woolf and her group. The article begins by discussing some of the potential advantages of AI-based tutoring systems. Intelligent technologies, the authors argue, may be able to customize themselves to the needs of individual learners. In addition, AI applications can also make asynchronous learning effective... . AI techniques will enable students to learn about selected material at their own rate before it is taught or as a supplement to course activities. Such systems might become routine supporting group collaborations of students-at-a-distance, exploration of hypothetical worlds, and the making and testing of hypotheses. Learning need not be constrained to place and time. (p. 100) After arguing for the desirability of AI-based technologies, the authors then tackle the question of what sorts of machines qualify as "intelligent," and able to produce these benefits. First, a machine might be intelligent if it is "generative," that is, if it independently generates instructional materials, problems, and hints appropriate to the student. Second, an intelligent machine might make use of modeling. It can model either the student-user, assessing the current state of knowledge and doing something based on the model, or it can use expert models of a particular knowledge domain. Third, an intelligent machine can perform "instructional modeling," that is, it can tailor the instruction to the changing level of the learner. Next, a machine is intelligent if it allows for "mixed initiative" interaction—a smart machine can initiate needed interaction with a student and can respond appropriately to the student’s action. Finally, a machine might qualify as "intelligent" if it is self-improving, and thus able to gradually improve its own teaching performance. While eventually AI-based tutors may have all of these abilities, currently most are much more limited. The authors, however, offer case studies of machines that are pushing the boundaries. They discuss the Cardiac Tutor, and various mathematics and engineering tutors. These applications are generative, they perform student and expert modeling, and, in the case of the mathematics tutors, are self-improving. The authors conclude their essay with an evaluation of these intelligent tutoring systems. They write, "Formal evaluation of student performance with intelligent tutors has shown that tutors can achieve performance improvement that approaches the improvement by one-on-one human tutoring compared to classroom teaching" (p. 138). Kenneth Koedinger discusses the famous Pump Algebra Tutor (PAT) in his article "Cognitive Tutors as Modeling Tools and Instructional Models." One focus of PAT is helping students develop multiple representations of algebra problems. With PAT, students can create tabular, graphical, and symbolic models of problems. The cognitive tutor chimes in as needed with just-in-time feedback. The tutor highlights possible errors with outline text—errors that the software recognizes by comparing student actions with its database of common errors. The student can also request a "context sensitive" hint. The tutor tracks the learner's progress on a problem and then, in a fairly complex process, selects the appropriate advice: The tutor chooses the hint message by using the production system to identify the set of possible next strategic decisions and ultimate external actions. It chooses among these based on the student's current focus of activity, what tool and interface objects the student has selected, the overall status of the student's solution, and internal knowledge of relative utility of alternative strategies. Successive levels of assistance are provided in order to maximize the students' opportunities to construct or generate knowledge of their own. (p. 158) PAT also supports learning through "knowledge tracing," which tracks a student's problem solving skills, identifies areas of difficulty, and presents problems in areas that have not yet been mastered. PAT seems to work well: field studies show a 15-25 percent difference between those classes that used PAT versus control groups. The most interesting part of the essay, however, is the final section in which Koedinger discusses PAT as a vehicle for teacher change. PAT offers a student-centered model of learning, that teachers, once they are exposed to it, often tend to replicate in other areas of instruction. Koedinger observes that, sometimes, teachers begin to borrow the problems, the representational tools, and feedback strategies embedded in the cognitive tutors. He writes, "Teachers began to use PAT problems in their regular classes and began to experiment with more student-centered learning by doing outside of the computer lab" (p. 165). The computer becomes a pedagogical role model. Chapter 6 deals with using intelligent tutors in reading education. In their essay, "Evaluating Tutors that Listen: An Overview of Project LISTEN," Jack Mostow and Gregory Aist discuss the development of computer tutors that listen as students read and that offer corrections if they hear students make mistakes—an impressive feat of technical engineering. Sometimes the tutor helps with only problematic words: the tutor highlights the word, asks the student to read the word, and then echoes the correct word after the child's response. At other times, the tutor rereads the sentence and lets the student try again. The tutor then listens to the student and offers additional feedback. Later versions of the tutors possess still more options. Students can, for example, listen to themselves reading a passage by selecting the "play back" option. The tutor can decompose a word (sounding it out while highlighting the appropriate letters) or offer rhymes for problematic words (e.g., if there is a problem with the word "wheat" the tutor may suggest that it "sounds like feat"). Finally, the authors discuss how the computer tutor allows them to do "invisible experiments." The tutor can vary its tactics in different situations and keep track of the results, thus creating some idea of what interventions are most helpful without disrupting the learning process with more intrusive research. The authors conclude their essay by pointing to empirical evidence suggesting that the listening tutor helps students read more effectively. In the next contribution to the volume, "Articulate Software for Science and Engineering Education," Kenneth Forbus attempts to show how articulate virtual laboratories (AVLs) and active illustrations can increase technical skill and general scientific literacy. Forbus's goal is to increase students' knowledge of qualitative physics. Qualitative physics deals with the conceptual underpinnings of the natural world, or as Frobus puts it, questions dealings with "what happens, when does it happen, what affects it, and what does it affect" (p. 236). One way to increase knowledge of qualitative physics is to experiment with relationships, but in some domains, like engineering thermodynamics, this is difficult and dangerous. Virtual laboratories like Forbus's CyclePad can perhaps solve this problem. In CyclePad, engineering students can manipulate a variety of parts (compressors, turbines, pumps, etc.) in a thermodynamic system. As a student designs a thermodynamic system, the program derives consequences from the design. In designing a system, students can ask questions that link back to an explanation database, ask to see diagrams of the systems they are developing, or calculate the economic costs. Also included is an on-board coach that works from a database of common contradictions, providing analysis and advice. The goal of CyclePad is to help students to get a clearer picture of the consequences of design decisions in a realistic environment. Active illustrations offer another technique of helping to grasp relationships between scientific concepts. As an example of an active illustration, Forbus discusses the Evaporation Laboratory. In this active illustration, students can see how changes in an environment, temperature, and container can change the rate of the evaporation of water, thus making more feasible experiments that would take a long time to do in the real world, and that would require travel from one climate to the next. In Chapter 8, "Animated Pedagogical Agents in Knowledge-Based Learning Environments," James C. Lester and his group describe their work that attempts to exploit the emotional connection we seem to feel to computerized characters and alter egos. They write, Because of the immediate and deep affinity that children seem to develop for these characters, the potential pedagogical benefits they provide are perhaps even exceeded by their motivational benefits. By creating the illusion of life, lifelike computer characters may significantly increase the time that children seek to spend with educational software, and recent advances in affordable graphic hardware are beginning to make the widespread distribution of realtime animation technology a reality. (p. 269) While the stated goal of "significantly increas[ing] the time children seek to spend with educational software" is certainly odd, it reveals a faith that experience with educational software can lead to learning. After discussing the reasons why animated pedagogical agents may be effective in motivating students, the authors discuss several of the agents they have developed including Herman the Bug, who helps students design plants and trees appropriate for given environments, Cosmo, a "helpful antenna-bearing creature with a hint of a British accent" that helps students think through computer network routing mechanisms, and Whizlow, an explanatory lifelike avatar that offers help in the domain of computer architecture. These agents serve both to represent the student in the virtual world and to offer advice when needed. When the student pauses for a long time or makes a mistake, the animated agents step in to offer appropriate help. The authors offer some details relating to how the agent is programmed to act in a smooth and helpful way, and also describe how the animated agents determine whether students have made mistakes and how the agents determine the correct response. Sometimes, a student misconception is exposed when a model of the student's past behavior is compared to a database of common novice misconceptions. Once the misconception has been identified, the appropriate response to that misconception is analyzed as it relates to the particular student's current context, and an appropriate agent response is generated. If the student continues having difficulty acting successfully, the agent can model the appropriate actions. The authors conclude their essay with a summary of some research evaluating the effectiveness of animated pedagogical agents. Chapter 9 is entitled "Exploiting Model-Based Reasoning in Educational Systems: Illuminating the Learner Model Problem." In this essay, Kees de Koning and Bert Bredeweg attempt to apply what is known about model-based reasoning to the design of educational software. The goal is to develop applications with deep subject matter knowledge that can interact intelligently with a learner. The interaction with the software is individualized as the program builds a model of the learner and compares it to a normative model in the subject matter domain; where the behavior deviates from the norm, the software develops an appropriate diagnosis and applies an appropriate pedagogical intervention. Now, in most other intelligent learning applications, the diagnosis of the problem is based on elaborate, detailed, and expensive "bug catalogues," which try to take all known ways in which people make mistakes in a given domain, and then compare the items in the catalogue to the learner's mistake. Conversely, de Koning and Bredeweg want to base the diagnosis on models that seek to "identify those sets of primitive reasoning steps whose correct behavior is inconsistent with the observations" (p. 308). There is no attempt to build a learner's mental fault model here; the diagnosis is only at the level of problem solving behavior. After discussing in some technical detail how the subject matter model is generated (the primitive steps of reasoning are generated on the basis of qualitative simulation models) and how the behavior diagnosis is created, the authors then describe their STARlight educational application which asks students to predict what will happen in an example situation from physics. When learners make a false prediction, the software probes their knowledge with several multiple choice questions designed to find out where the exact problem is. Is this probing effective? During a small trial run, the program was often successful in finding the error and correcting it (or, more often, the learners realized their own mistakes during the probing process). It could be said, then, that in this application, the diagnostic instrument fruitfully drives the dialogue with the learner. The "help" option does not stand apart from the application, only to be called upon in rare circumstances; rather, the application is built around finding and correcting errors. The next two chapters are different from the rest of the collection in that they do not discuss particular technologies at all; rather, they discuss ideas about learning that may be included in the development of technological artifacts. The first of these two articles, "The Case of Considering Cultural Entailments and Genres of Attachment in the Design of Educational Technologies," by Lisa M. Bouillion and Louis M. Gomez, discusses cultural considerations that will be necessary to consider when technologies actually enter classrooms. They argue that, as a function of design, artifacts come with a set of cultural entailments, representing the goals, expectations, histories, values, and practices associated with a particular community. Left unquestioned these entailments may clash with the entailments of a community targeted for use of the innovation. (p. 347) As an example of this clash, they investigate the case of a particular educational innovation designed to break down the walls between the school and society. This innovation attempted to involve students in "real world" problems faced by community institutions. Community groups were approached and asked what real help they could receive from children, and the students were then to be involved in filling this need. In a suburban school, the cultural entailments of this innovation, which included preparing students to contribute to an established community, meshed well with the culture of the students. The students were able to contribute to an awareness campaign on behalf of United Way. In an inner city school, however, the entailments of the innovation did not mesh with those of a more suspicious minority community. In this community, the innovational entailments would need to change from fitting in to the established community to changing it. For this reason, the inner-city teachers suggested that their students work on their own projects, and this eventually became an activist movement to clean part of the Chicago River. The innovation was, then, molded to the requirements of the community. The authors argue that this negotiation between educational innovation and cultural entailments must always take place, a fact that must be incorporated into the design of educational artifacts and technologies. It would have been interesting to see the authors use this framework to examine the technologies described in the rest of the book. The second article to discuss considerations that should be used in designing technologies is that of Paul J. Feltovich, Richard L. Coulson, and Rand J. Spiro: "Learners' (Mis) Understanding of Important and Difficult Concepts: A Challenge to Smart Machines in Education." In this contribution to the volume, the authors explore the question of why some misconceptions about the world are so hard to change, while others are not. Entrenched or "stable" misconceptions, they argue, tend to be simpler and more easily understood than their more correct counterparts. The elements that make up an entrenched misconception tend to fit together nicely and in such a way so as to provide mutual support. An entrenched misconception "is also more concrete, carries with it more salient examples and analogies, and is more congruent with the intuition" (p. 366). The authors take a notorious example from medical education relating to blood circulation, show how the example manifests the features of being a stable misconception, and construct a list of "knowledge shields" that allow students to evade changing their minds when faced with arguments showing the old idea to be false. The authors end their article by arguing that the content most filled with these sorts of misconceptions are those areas of knowledge that are "the continuous, the simultaneous, the interactive, [and] the conditional" (p. 375). The authors think that this type of subject matter tends to be neglected by educational technologies, including, with a few exceptions, the smart machines described in the present volume. Kirstie L. Bellman's contribution to the book, "Building the Right Stuff: Some Reflections on the CAETI Program and the Challenge of Educational Technology," is a retrospective look at the computer-aided education and training initiative (CAETI), one of the most important R&D programs in educational technology ever sponsored by the government. In the article, Bellman gives a brief history of CAETI research. She also discusses the program's technology insertion experiments in the Department of Defense's K-12 schools for overseas dependents. The development of technology for education is difficult, she argues, because it requires competence in a wide variety of fields, for example, in academic subject matter, learning processes, cultural and small group processes, and engineering and product design. Educational technology implementation is also difficult because it requires a "tedious conversation" among the parents and students and all those who develop, use, fund, and research educational technology. Bellman's essay, I believe, does a fine job giving the reader the flavor of these conversations. CAETI's tedious conversations took place within and among several clusters of projects. CAETI aimed at developing (1) "smart navigators to access and integrate resources" (SNAIR) to help students retrieve and use appropriate online information, (2) "collaborative applications for project-based educational resources" (CAPER) to create MUD-like spaces for intellectual interaction, and (3) "expert associates to guide individualized learning" (EAGIL) to devise intelligent tutors, some of which are described in other chapters of the book (e.g., BELVEDERE and the PUMP algebra tutors). The initial idea was that the different participants—each working on a separate project—would eventually work together and combine their applications to create truly complex and subtle applications. It was not feasible, however, given the vast differences among the stakeholders and levels of commitment, that this cooperation be forced from the top down. Thus, it was only required of the participants that they make their programs open (that they "grow ears and a nervous system"). With this minimal compliance standard, it was hoped that curiosity among participants would take over ó an organizational strategy that has apparently fostered many fruitful collaborations. Indeed, after discussing the difficult issues involved in evaluating the products CAETI eventually produced, Bellman spells out what she thinks were CAETI's successes and failures. With regard to the successes, Bellman mentions the collaboration that was fostered. She also counts as a success many aspects of CAETI's technology insertion program: Many of our decisions were sound: don't throw the technology over the wall, use methods like TIP [a technology insertion plan which works with the teacher to combine the technology with the teacher's curriculum objectives and instruction and assessment strategies], sponsor a diversity of subjects and grades, and use technology to support a variety of learner and teacher styles. We also made good decisions with regard to teacher training and involvement. We worked with the unions to make sure that the teachers foray into technology was a safe experiment. We tried to develop in each school a critical mass, build up mentoring and advocates. . . .We also made the decision to pay the teachers in at least some fashion for their extra time and costs to train. We also made the good decision to start with teachers who were interested and enthusiastic. (p. 408) With regard to the things that should have been done differently, Bellman writes that CAETI did not take full advantage of students in both running the technology infrastructure and having input on developing their own educational experience. Bellman also regrets not taking fuller advantage of wireless technology (including wearable and mobile devices like palm pilots). She ends her essay by pointing to the challenges that will face subsequent initiatives aimed at developing educational technology: continuing the momentum from one program to the next, helping teachers become active researchers in the model of medical personnel, and facing the ever-present questions of the aims of education (one of the biggest problems in development technology is deciding what, exactly, should be the point of the technology). Comments and Questions The tone of this book is generally more guarded than in many books about educational technology. True, the authors are all optimistic about the possibility of technology revolutionizing education, but there is little of the inflated rhetoric that can so often accompany books about educational technology. The authors are generally aware of the difficulties that come when new technology enters the classroom, and many of the essays are stories of the attempts to overcome these obstacles. The tone of the articles thus tends to be somewhat "older and wiser" than other books in the genre, and this experience shows. For example, many of the authors admit that people who develop smart machines should listen to teachers, students, and parents as much as they need to talk (see pages 163, 231, 408). Compare this to the rhetoric of other technology advocates, like Don Tapscott (1999), who see teachers as more of an impediment than anything to the advance of the technological steamroller. Effective technology insertion is much more than just throwing the new machines "over the wall" into a bewildered school that never asked for them. The authors seem to realize this. The authors also seem generally motivated to provide real answers the question of why their technologies are needed. A skeptic can and should ask, with every technological innovation in education, if the new innovation is really necessary, or if it is just more technology for technology's sake. As school closets fill up with barely-used and now obsolete computers, these questions press even harder: Is a new educational technology really filling a need, or is it just a way for developers to express their technical virtuosity or for corporations to line their pockets? Why is the technology is necessary and what it can do that a teacher (or even a relatively untrained TA) cannot do? Many of the authors try, with varying degrees of success, to answer these questions. Some have tried to justify advanced technology in education by simply invoking the "economics of scale"—greater numbers of students can be served through these smart machines for a lesser cost (i.e., it is hoped that technology will reduce educational labor costs). Invoking the economics of scale, however, is often suspect when it comes to educational technology. The ever changing infrastructure, the support staff, and the applications themselves are not cheap—far from it. Often, when an institution reaches the point it could start saving money on a technological implementation (the so called "cross-over point"), both the hardware and software have become obsolete. In this and in many other ways, the rapid development of technology is the worst enemy of actually implementing technology in education. There is no time to perfect it, nor is there any time to save money with it, before the next new thing comes along. If technology rarely delivers on its promise to save money beyond human teachers, perhaps there are pedagogical benefits that make the technologies worthwhile in spite of the cost. And indeed, there are some powerful pedagogical arguments presented for smart machines in the book. For example, computers allow for the "invisible experiments" of the type described by Mostow and Aist, allowing an educator to carefully track pedagogical effectiveness in inconspicuous ways beyond what any teacher could previously do. This, together with the idea of "automated experiments," could revolutionize educational research, while at the same time raising ethical concerns about the constant surveillance such technology makes possible. Another example is that offered by de Koning and Bredeweg, whose application traces errors in a student's problem solving process to the basic reasoning steps that could not have been applied correctly. This process is so intricate, they argue, it would be "difficult or even impossible for human teachers" (p. 326). This claim is debatable, of course, but if true it would serve as a justification for the smart machine in its domain. In addition, some of the contributors argue that their smart machines create a world in which students are not afraid to fail (see, for example, p. 59). The computer does not laugh at you, look exasperated, or make snide remarks. The very social decontextualization, bemoaned by some technological critics, makes possible an environment where it is now tolerable to learn from mistakes. Granted, the critic could still ask when students would learn to deal with situations where social pressure exists to perform. As Quintilian pointed out long ago, schools are in a public setting because we want students to learn how to act in a public sphere, with all the pressure to perform: “let him who . . . must live among the greatest publicity, and in the full daylight of public affairs, accustom himself, from his boyhood, not to be ashamed at the sight or men” (1965, p. 22). But this ability to let students make mistakes freely, away from the threat of social scorn, could still turn out to be one of the most important educational possibilities these new smart machines offer. This book shows how the new generation of educational technology has made the student-computer interaction much richer. There is much more to these programs than simply typing in a numerical answer and being told whether the answer is right or wrong. I must admit, though, I still found some of the interaction to be quite shallow. The applications described in the book can only handle a given number of pre-set responses. If a student says something off-the-wall or otherwise unexpected, the programs would be unable to respond. When the simulations and tutors described in the book are trying to teach a clearly rule-governed procedure, they do quite well. But there is still no smart machine, apparently, that can help students to think creatively, or otherwise outside of rule-governed frameworks. A tutor can be developed to help students solve math problems, but not to help them become artists. Science, which can be alternately rule-governed and creative, in the hands of these smart machines, generally reduces to only a rule governed procedure. None of the scientific simulations allow students to develop their own research questions in depth, and even in the research questions that are given, there seems to be little room to develop one's own hypothesis if it is not a pre-set option. This problem of interaction is compounded because many of the applications described in the book rely on comparing expert models with student-generated models. This method of using expert models reveals both the brilliance and baggage of bringing AI into educational technology. Success is defined such applications as living up to the expert model. While much educational mileage can be traveled on this road, it does not lend itself to developing creative, revolutionary thought. If a thought is not within the expert model, it is deemed a failure. The problem, then, even in these smart machines, is that students are treated as rule-following machines—the better they follow the rules, the more they will replicate what an expert would do, and the more successful their computer interactions become. In 1986, Michael Streibel argued that computerized tutorials demand that human beings act as data-based, rule-following, symbol-manipulating, information processors. By demanding pre-set responses, creating a narrow model of the learner (as inputter of data and not as a full human being), and comparing this model to expert standards, the educational discourse contracts too tightly. He writes: Uniformity in education is enforced not only because the instructional system attempts to shape a uniform product (i.e., pre-specified learning outcomes) but also because the very conceptualization of the individual places semantic and syntactical constraints on acceptable language for the discussion of human beings. This in turn makes it impossible to express and legitimize other conceptions of human beings, educational goals, and methods outside of the technological framework. (p. 150-1) I do not see that, some 15 years later, we have come any closer to making the computer/human discourse any richer. Although it is now dressed up in fancier clothes, your AI date still doesn't know how to carry on a very good conversation. I have other worries about many of the applications described in the book. One of the best things about simulations, for example, is that we can seem to speed up the natural processes of the world. For example, the learner does not have to wait for days or travel vast distances to test the difference in the evaporation rates in Forbus's Evaporation Laboratory. The simulations break down the walls of time and space and permit an inquiry that would otherwise be impossible. It is important to realize, however, what this speeded-up world does not teach about the process of inquiry. It does not teach students to be patient, for example, and to wait for data that may be difficult and pain-staking to accumulate. Moreover, in their fear of student boredom, the simulations often become interlaced with flash and glamour. The science simulation Wolf, for instance, discussed by Schank and Neaman, contains exciting sequences of scientists leaning from helicopters, tranquilizing and examining a large wolf. It does not have scenes of the scientist reading quietly in the library for hours on end, or playing the political cards necessary to get government funding. The exciting and glamorous tend to be emphasized, and thus the simulation may give a false impression about how inquiry actually proceeds. Some of the authors write about trying to imitate the excitement and passion of video games (see, for example, pages 97 and 266) in their educational applications. I worry that this is being a bit disingenuous: real inquiry is often quite humdrum and labored. Although I worry about some of these points, most can be answered by simply adding other types of experiences ("real-life" dialogue to supplement the shallow computer interaction and “real-world” inquiry that requires the development of scholarly patience) to those of the computer tutors and simulations. To the extent that these additions are made, my worries will largely be unfounded. One especially impressive aspect of the book is the way it reveals an important reason for taking educational technology seriously, namely, developing educational technology demands that we continually rethink educational questions. As we develop and evaluate educational technologies, old questions are infused with new life: What is the relationship between a teacher and student? What should happen when students make an error? What is the best way to teach math? How do students learn? What is the purpose of education? As the authors wrestle with their technologies, they do some fine thinking about questions such as these. For example, some of the pure educational thinking in this book relating to student mistakes and misconceptions was exceptional (see Schank and Neaman, pp. 59-66, de Koning and Bredeweg, pp. 319-326, Feltovich, Coulson, and Spiro, pp. 349-375). Biswas and his group present a wonderful discussion of how learning-by-teaching may, or may not, be an effective strategy (pp. 79-84), and Lester's group present some interesting ideas about how we may learn through alter-egos (pp. 270-272). Perhaps this is the real legacy of the smart machines in education—we get smarter about education as we try to develop them. References Quintilian. (1965). On the Early Education of the Citizen-Orator. Indianapolis: The Bobb-Merrill Company. Streibel, M. (1986) . A Critical Analysis of the Use of Computers in Education. Education Communications and Technology Journal, 34, (3) 150. Tapscott, D. (1999). Growing up Digital: The Rise of the Net Generation. NY: McGraw-Hill Trade About the Reviewer Bryan R. Warnick is a doctoral student in Philosophy of Education at the University of Illinois at Urbana-Champaign. His research interests include philosophy of technology, ethics, and philosophy of mind as these disciplines relate to educational theory and practice. He holds a B.S. degree in Philosophy and Psychology from the University of Utah, and a M.A. in Philosophy of Education from the University of Illinois at Urbana Champaign.