Benefits Seen for Students Teaching Virtual Pupils

Educators have long held that peer tutoring can help students learn, and emerging research on students working with computer characters points to one possible reason why: Teaching begets learning for the teachers.

Researchers at Stanford University’s AAA Lab and Vanderbilt University’s Teachable Agents Group call it the “protégé effect,” which posits that students will work harder, reason better, and ultimately understand more by learning to teach someone else—even a virtual “teachable agent”—than they will when learning for themselves.

Being social is a frame of mind, said Daniel Schwartz, the director of the Palo Alto, Calif.-based AAA Lab, a social and cognitive learning research center. “These kids know these characters aren’t alive, but they get engaged with the narrative and play pretend, and it brings out a lot of good behaviors,” he said.

Most studies of peer-mediated learning and reciprocal teaching focus on improvements for the students being taught, rather than the advantages for the student-teacher. Teachable Agents researchers instead study how the act of teaching, both in students’ effort and reflection on thinking, improves their learning.

The research teams started talking about teachable agents nearly a decade ago, in response to too much student dependency on answer feedback in virtual tutoring programs. Rather than understanding a concept, a student tended to try answers randomly until one worked. “We wanted to build systems that could help students learn with understanding,” said Gautam Biswas, a computer science professor and head of the Vanderbilt Teaching Agents Project.

Building Virtual Betty

The teams since have developed a half dozen different computer programs in which students customize a virtual agent and teach it mathematics or science concepts. The agent questions, misunderstands, and otherwise learns realistically. For example, in the Betty’s Brain program, a concept map represents the character’s thought process, and students teach her by making logic chains. Showing how garbage contributes to global warming, for example, requires connecting eight separate causal inferences. The student can test his or her agent’s knowledge, ask it to explain its answers, and correct misunderstandings by adding new information to the map.

So far, all of the experiments have been relatively small-scale—a few schools or grades at a time—but they have explored dozens of permutations of how students from kindergarten through college interact with computer agents.

In a series on 5th, 6th, and 8th graders, researchers used the same software, curriculum kit, and science teachers, but told one group of students that they would be teaching the characters and the other group of students that their characters were avatars, representing the students. The student-teachers spent nearly twice as long reading the study materials, and more time checking and re-editing the concept maps, while the students learning for themselves spent more time in an online chat with other students and playing the quiz game. Moreover, experiments showed that the student-teachers were more likely than the learners to voice concern and responsibility when the characters answered quiz questions incorrectly, and also were more likely to study and do other work after the first quiz to improve for the next one. In the end, the student-teachers performed better than learners on both the informal quiz shows and a separate post-test.

“When you are looking at something yourself, you can fool yourself into thinking you know everything, but when you have to communicate it to someone else, you realize that you’re really not being precise enough,” Mr. Schwartz said.

Unlike real students, the agent always explains how it came to an answer, and in separate studies the researchers have found student-teachers adopt their agent’s thought process. Over three science units, student-teachers became better than contol-group learners at making short, medium, and long lines of reasoning.

More recently, Mr. Biswas added a mentor agent, “Mr. Davis,” to help student-teachers learn new reasoning strategies to teach their characters. The classroom teachers also receive feedback on students’ common mistakes, so that they can discuss problem areas outside the virtual environment.

“Teaching seems to have a positive effect on learning and on top of that, giving meta-cognitive feedback also improves learning,” Mr. Biswas said.

Both labs are moving to bring the lessons from virtual teaching to flesh-and-blood classrooms. This year, the Vanderbilt lab is exploring ways to help students transfer reasoning strategies to other subjects through schoolwide projects, while the AAA Lab has started to compare the protégé effect for students teaching both real and virtual students; studies from both labs are due in coming months.

Douglas H. Fuchs, a special education professor at Vanderbilt and developer of Peer-Assisted Learning Strategies, a popular reciprocal-learning program, said he has seen similar benefits in his own studies of real-life students involved in peer tutoring in math. In those studies, high-achieving students, as well as tutees, benefited, which Mr. Fuchs said could mean “there really is something important for the ‘teacher’ if the context is smartly set up and children are provided with appropriate training, guidance [and] direction.”