Power Learning Project

Power Learning: Using Computers as Teaching Machines

Latanya Sweeney

Laboratory for Computer Science
Massachusetts Institute of Technology

Links within this paper:


We conducted a study in which students learned part of their computer programming material using the traditional lecture, book and assignments, while others had the additional benefit of a computer tutorial that was responsible for teaching a small segment of the course material. Operational definitions for teaching machines, computerized teaching machines and power learning are provided. This study identifies student attitudes towards a tightly-structured computer-assisted instructional program. The computer was responsible for teaching and seamlessly incorporating practice into the session. In the study some students received traditional training (lecture, book, and assignments), while other students received the additional use of a computerized teaching machine in the teaching-learning of the C programming language. For students using the teaching machine, the computer constantly engaged students in reinforcing and interactive activities. Each student worked at their own pace, and at a time that best suited them. The computerized teaching machine spent more time on the subject matter than could be afforded during class time, and gave each student individualized attention. All students were then given an aggressive problem to solve based on their newly acquired knowledge. Student attitudes revealed an intense desire to include the teaching machine into the course structure and rated the teaching machine as useful as lecture and much more effective then the textbook.

Power Learning: Using Computers as Teaching Machines

The use of computers as teaching machines is not a new idea. Over 55 years ago, Sidney Pressey, in the wake of the industrial revolution in education, described in 1926 a machine that "test and also teaches. " B. F. Skinner, not knowing of Pressey's work, demonstrated a machine to teach arithmetic in 1954 (Skinner, 1987).

The widespread availability of computer technology offers an untapped resource in education. Researchers in this area have named such uses of computers as computer-assisted instruction (CAI) and computer-based education (CBE). The most popular example of computer-assisted instruction is Seymour Papert's LOGO, which engages the computer as a teaching-learning tool. LOGO, based on a Piagetian model of learning, proved to captivate and educate children by placing the child amidst a mathematical programming world in which the student writes programs to control the movement of a turtle on the computer screen. Problem solving skills, geometry skills, and programming skills improved (Brookline LOGO Project, 1979; Papert, 1980).

Contrary to the teaching structure proposed by Skinner and Pressey, LOGO represented an era in American educational psychology, where students were no longer told things, but an environment was created in which the student could discover things for themselves. Memorization was replaced with discovery. As an educational pedagogy, LOGO is the dialectical opposite of Skinner's teaching machine. In LOGO, as with most simulations, the computer provides an environment in which the student learns by doing and exploring. The only structure is that imposed by the human teacher through homework and class assignments. The success and educational merits of LOGO must be underscored. As the old adage states, "we do not want to throw the baby out with the bath water." The choice is not simulations versus teaching machines, but the proper use and implementation of both. Our focus in this study is to properly define and evaluate the role of computerized teaching machines.

Skinner makes an arguement that showing and telling are ways of priming behavior. "We learn when what we do has reinforcing consequences. To teach is to arrange such consequences," and by carefully "constructing certain contingencies of reinforcement, it is possible to change behavior quickly and maintain it in strength for long periods of time" (Skinner, 1987). The central process is called operant conditioning. A growing body of research in computer-assisted instruction supports Skiner's view. A structured teaching environment holds a lot of potential and promise for the computer to provide meaningful instruction. (Siegel and Davis, 1986).

Unfortunately, the vast majority of educational software currently available is neither a simulation nor a structured teaching environment. In fact, these programs offer nothing more than drill over a small and limited domain. Common computer tutorials simply display the contents of a textbook on the screen, with immediate search and index capabilities and often some animation. Yet, the responsibility for teaching, the integration of practice, and the assurance of mastery is not within the scope of these programs.

The biggest problems hindering work on teaching machines are: (1) the vast amount of development time required to produce properly structured programs; and, (2) the lack of concrete programming guidelines to insure the student's mastery. The guidelines for providing a meaningful teaching-learning environment on the computer must be analogous to the human teaching experience.

In this study we designed and built a well-structured teaching machine covering a small topic of course material. After establishing programming guidelines that seamlessly incorporate instruction and practice and assure the student's mastery of the material, we named such guidelines power learning. Some students used power learning while others did not. Student attitudes towards the class, lecture, the textbook, and the computerized teaching machine were measured for both groups.


Operational definitions

Teaching machine. A teaching machine is a non-human device which is responsible for the carefully structured presentation of new material and the seamless incorporation of drill and practice into the learning process. A teaching machine is an aid to teachers, but is not the same as traditional aids, such as a chalk board or a book. In traditional teaching aids, the teacher remains responsible for the presentation of material and the choice of when and how to incorporate questions, drill and practice. In a teaching machine, these operations are performed by the machine.

Computerized teaching machine. A computerized teaching machine is a teaching machine based on the use of computer technology to produce the machinery used to teach. This will involve some form of a stored, pre-programmed sequence of events. It is important to emphasize that not all computer-assisted instruction can be considered as teaching machines, since simulations such as LOGO, and programs that only engage the student in drill and practice, are not responsible for the structured presentation of material.

Power learning approach. Without a lengthy discussion of the guidelines we follow, power learning can be defined as a tightly-structured, step-by-step learning environment, that can guarantee each student's mastery of the material presented in the environment. Such an environment must first adhere to these fundamental requirements:

  1. each step must be simple;
  2. the communication of ideas and concepts at each step, must plant the same, unambiguous, specific model in the mind of each student;
  3. the next step in the instruction sequence must be so incremental as to ensure each student's ability to intellectually take the step; and,
  4. the computer must constantly incorporate practice and review within the instruction.
The resulting approach is always a sequence of very simple, incremental instructions. Each step is so simple that a student can not help but understand what is presented. The activity may even seem trivial to the student. Yet, when a student completes a set of these steps, the student gains solid understanding of a complete subject matter. A more graphical way of saying the same thing, is that by taking many very small steps, the student ends up traveling a long distance.


The subjects were 34 undergraduate and graduate students currently enrolled in a first semester computer programming class at Harvard University. The computer language being taught was the C programming language. No previous experience in computer programming was required. The participants volunteered from a total class of 90 students. Thirteen of the students used the power learning program, and 21 did not. Eleven students completed a pre-test survey. Of these 11 students, 8 students used the power learning program. All participants completed a post-test survey.


The materials consisted of the power learning software program and computers, a pre-test, a post-test, lecture, sections, additional course handouts, a textbook and a homework assignment.

Power learning software program. The power learning program built for this study covered the teaching and learning of the random number generator and Monte Carlo simulations in the C programming language. The program was implemented on 3 computer platforms --IBM PC and compatibles, Macintosh computers and Unix computers. Participants could use either of these versions, as many times as desired, at home or at school, at anytime of day or night that best suited their schedule. The development time ran 40 minutes of development for each minute of use. The total run time was 45 minutes for most students. So, the total time to develop the program was 30 hours.

Pre-test and post-test. Some students were given a pre-test survey. All students were given a post-test survey. The pre-test consisted of five, 8.5 x 11 in. sheets of papers containing a total of 46 questions. The post-test was identical to the pre-test, except it contained an additional cover page of 5 questions that identified the student as using or not using power learning, and completing or not completing the pre-test. A copy of the post-test is included as Appendix A. The questions were grouped in 4 major groups: (1) feelings about the class and personal performance; (2) attitudes about the usefulness and practicality of using power learning; (3) a survey of the students knowledge of the material being taught; and, (4) a table of attributes to describe lectures, sections, the textbook, and power learning.

Almost all questions had a continuum line, ranging from "low" to "high." Students placed a mark on the continuum denoting their response. The table of attributes had discrete ratings of low (1) to high (5). The attributes queried were: fun, challenging, unfair, difficult, and frustrating. Weiner (1980) has shown that "thoughts determine what we feel and feelings determine what we do." So asking students about their feelings offers a good prediction of their action.

Design and Procedure

Students volunteered to participate in the study. After completing the pre-test questionnaire, some students were not allowed to use the tutorial immediately. We explained to these students that they constituted the control group. Of these students 90% verbally protested and immediately insisted on one-on-one discussions with me. These students illustrated extreme agitation, including crying. This made it clear that the students volunteering for the study were very pressured and feared not being able to complete the assignment without as much assistance as possible. The next day all students were allowed access to the power learning program.

An aggressive homework assignment was given based in part on the material covered in the tutorial, but demanded mastery of other new material as well. The assignment required the creative use of the random number generator to produce a graphical simulation of a drunk person walking down a street. The students had to track the number of steps and be able to change the width of the street. The students had one week to learn the material and complete the assignment. During that time, they had one lecture, one section, course handouts and the textbook available. Each student was expected to work alone, although tutors were available to offer assistance.

At end of the week, the homework was collected, and both the students using the tutorial and those who did not, completed post-test questionnaires.


The results reported here are based on the median values of the composite scores. No significant difference appeared between the medians and the averages on each question.

Eleven of 13 students (84%) stated that power learning was 33% better than expected. One student stated power learning was exactly as expected and one student stated it was far worse (citing an expectation for power learning to cover all the material needed for the assignment).

The students using power learning spent 17% more time with the package than they expected.

Both students using power learning and those who did not, believed power learning could help their performance in class. But using the program, increased the belief in the usefulness of power learning by 10%. Further, when asked how much they could learn from power learning in the pre-test, all students expected a 30% increase in learning; but students who actually used the program reported a 60% expectation afterwards.

The students participating in the study represented a cross-section of the students: 12% of the students participating in the study expected a failing grade in the class; 32% expected an average grade, and 56% expected an above-average grade. The range of expected grades ranged from 8/100 to 100/100. So, it was not the students with the lower grades that flocked to the study. Also, no significant difference was found in student attitudes towards power learning based on their ranking in the class.

The students identified their ability and their self-confidence, as it related to the course, at 75% of the maximum possible. Further, students characterized the class as requiring high effort, 88% of the maximum possible; humility at 50% of maximum possible. Other attributes such as anger and guilt rated lower at 25-33% of the maximum possible. These findings are consistent with Jagacinski and Nichols (1987) showing that the class environment requires high effort with low success, and some students, attribute their inability to succeed on external factors. Further, because humility lies at 50%, the class consists of a mix of ego-involved and task-involved students.

As for the specific material taught by power learning, students who used the power learning software, reported learning 78% more of the material than the students who did not use power learning. However, this is still 18% below what the students expected to learn about the material beforehand.

Attributes assigned to power learning rated fun as expected, which was the same as lecture and section, and twice as fun as the textbook. Power learning was half as challenging as expected, which made power learning half as challenging as lecture, section and the textbook.

Students who did not use the power learning software expected power learning to be 3 times more unfair than lecture, section and the textbook, but students who used power learning found it to be just as fair. Likewise, power learning was expected to be as difficult as lecture, section and the textbook, but was found to be 33% less difficult than the others. Finally, along the same lines, power learning was expected to be as frustrating as lecture, section and the textbook, but was found to be half as frustrating, receiving the best rating possible. In fact, power learning received the lowest rating possible in unfairness, difficulty and frustration. The textbook rated worst, being assigned attributes as extremely unfair, difficult and frustrating. Students characterized the textbook as challenging, difficult and frustrating, while power learning was fun and far less demanding. Students described lecture and section as fun and challenging, but difficult.


At a time when even smaller class sizes have soared to 25 or 30 students, it is surprising that we do not employ common computer technology to personalize instruction. No one can deny that computerized teaching machines offer personal attention to students. Further, continuous monitoring of students is not typically available. Lessons are not personalized to move at a pace determined by the needs of each individual student. Yet, teaching machines provide exactly these features and can be implemented with current technology.

Power learning shares Skinner's belief that good instructional programs "maximize the effect of success as a conditioned reinforcer by asking students to take small steps and making every effort to help them do so successfully" (Skinner, 1987). Success has a powerful impact on the student, and power learning ensures success by the careful selection of incremental steps and the constant interaction with questions. Students get trapped in the power learning environment because paying attention has reinforcing properties.

Skinner believed that with the help of teaching machines and programmed instruction, "students could learn twice as much in the same time and with the same effort as in the standard classroom" (Skinner, 1987).

The results of the study show that students overwhelming preferred power learning to the textbook. Students who did not use the power learning software expected the tutorial to be fun and very challenging , but unfair, difficult and frustrating. The students who used power learning assigned the attributes: fun, not too challenging, very fair, easy, and not frustrating at all. At the same time, these students reported learning 60% of the material, while those students who did not use the program reported learning only 30% of the material. So, the incremental steps which leads to constant success allowed the students to use the program longer and yielded a friendly, supportive learning environment.

This study has shown that we can build a competent power learning system that presents the mechanics of the C language in small, manageable doses. But, this only accomplishes part of what is required in a programming class. Each student must also demonstrate mastery of the subject matter by constantly using newly acquired knowledge to solve problems.

It is quite easy and common for students to understand program presentations and concepts discussed in conversation, in a book, or even on a computer screen. But, when left alone to solve a slightly different problem, the student is lost and confused. Even mastery of how the C tools work, does not guarantee the student's success at putting the pieces together to solve problems.

To address this concern, first, power learning incorporates special activities that involve the student in applying newly learned material to specific problems. This way, the student not only gains mastery of a function or feature, but also recognizes the conditions under which to use this knowledge.

Second, power learning allows the student to learn from the mistakes of doing. The student builds programs under the supervision of power learning. This allows power learning to evaluate mistakes and make appropriate recommendations.

In using conceptual models to teach computer programming in the BASIC computer language, Bayman and Mayer (1988) identify the skills and types of knowledge required in learning how to program --syntactical, conceptual and strategic. Syntactic knowledge refers to the knowledge of facts about how programs are put together. This is analogous in writing English sentences, to knowing where to place commas, periods, and other punctuation and the word order within sentences. By conceptual knowledge, we refer to what is taking place inside the computer to make it behave as it does. You can't see the act itself, but you can measure the consequences. To continue our writing analogy, conceptual knowledge concerns the ability to abstract beyond the words to the bigger picture of what is happening in a story or paragraph. Finally, the ability to incorporate syntactical and conceptual knowledge to solve new problems, is termed strategic knowledge. In writing, these are the ways we've learned to use outlines and arrangements of information to put together essays, book reviews and the like. Given the array of skills the student must master in learning to program, the success of power learning in this domain predicts great success in other educational arenas as well.


Bayman, P. and Mayer, R. E. (1988). Using conceptual models to teach BASIC computer programming. Journal of Educational Psychology, 80, 291-298.

Brookline LOGO Project (1979) Final report of the Brookline LOGO Project. Cambridge: Massachusetts Institute of Technology, Artificial Intelligence Laboratory.

Jagacinski, C., & Nicholls, J. (1987). Competence and affect in task involvement and ego involvement: the impact of social comparison information. Journal of Educational Psychology, 79, 107-114.

Papert, S. (1980) Mindstorms: children, computers, and powerful ideas. New York: Basic Books.

Siegel, M. A. and Davis, D. M. (1986) Understanding Computer-Based Education. New York: Random House.

Skinner, B. F. (1987). Programmed Instruction Revisited. The Education Digest, 52, 12-16.

Weiner, B. (1980). May I borrow your class notes? An attributional analysis of judgments of help giving in an achievement-related context. Journal of Educational Psychology, 72, 676-681.

Appendix A

The following is a copy of the post-test given to students. The pre-test was identical except the first page was not included.

Copyright © 2011. President and Fellows Harvard University.   |  
IQSS   |    Data Privacy Lab   |    [info@dataprivacylab.org]