To appear in Educational Technology
We describe pedagogical agents called Teachable Agents (TAs) where students learn by
teaching the computer. We describe an example of a TA, and discuss the features that
allow students to capitalize on learning-by-teaching interactions. These include (i)
explicit well-structured shared visual representations, (ii) independent performance of
the agent, (iii) the agent’s ability to model productive learner behavior, and (iv)
embedding the agent in environments that support teaching. Finally, we describe new
directions for TAs including new models of homework practice, assessment, and video
game environments.
Almost everyone has had the experience of learning when teaching. Many
graduate students observe they never really understood a topic until they had to teach it.
To cultivate the benefits of learning by teaching, we have created a special kind of
pedagogical agent that we call a Teachable Agent (TA). Students teach their TA and then
assess its knowledge by asking it questions or by getting it to solve problems. The TA
uses artificial intelligence techniques to generate answers based on what it was taught.
Depending on the TA’s answer, students can revise their agents’ knowledge (and their
own).
TAs do not replace real students. But, they do provide unique opportunities to
optimize learning-by-teaching interactions. TAs, for example, always make their thinking
visible, something that not all students can do. This raises the question: What aspects of
learning-by-teaching can we maximize with TAs? We start with a concrete description of
a TA. We then describe four core learning-by-teaching design principles that we believe
can maximize learning-by-teaching. We conclude with instances of how these principles
enable us to introduce exciting new technologies that leverage learning-by-teaching.

BETTY: AN EXAMPLE OF A TEACHABLE AGENT
For this article, we spotlight our agent named Betty. Students teach Betty by
creating a network of entities and their relations, much like a concept map. Figure 1
provides an example. Students use a point-and-click editor to create nodes (e.g., LDL,
arterial plaque) and connect pairs of nodes with links. The links are labeled (e.g., LDL
builds up arterial plaque) and categorized using a pull-down menu (e.g., builds up implies
and increase) (Biswas et. al, 2005).
Figure 1. The Teachable Agent Betty. Students teach Betty by making a
concept map. Once Betty has been taught, she can answer questions by tracing links
through the concept map.
At any point, students can ask Betty a question to see how well she is learning.
Figure 1 shows how Betty animates her reasoning for the question, “What happens to
heart disease if exercise increases?” To make it easier for the student to follow her
reasoning process, Betty breaks down the explanation into parts. For example, Betty
reasons that exercise increases HDL cholesterol and that exercise decreases LDL.
Increasing HDL and decreasing LDL together result in decreased arterial plaque, which
in turn decreases the risk of heart disease. To complement her graphical thinking, Betty
unfolds her reasoning in text (lower panel). Students can observe Betty’s conclusions

and decide whether they need to revise what they have taught Betty. Betty can also take
a quiz composed by a classroom instructor but automatically scored by the computer. So,
instead of students taking the quiz, they can watch their agent perform and receive
projective feedback on their own knowledge.
FOUR CORE PRINCIPLES OF TEACHABLE AGENTS
Betty is one instance of a Teachable Agent (for other instances, see Schwartz et.
al.; in press; Blair and Schwartz, 2004). Like all our TAs, Betty includes features that we
think enhance the experience of learning-by-teaching. Some features are implicit in the
TA metaphor. For example, unlike a human learner, a computer agent will not be hurt if
the student is a bad teacher or has missing or incomplete knowledge. Other features
derive from learning-by-teaching design principles that we have tested with hundreds of
students of all ages (for empirical reviews see, Biswas et al., 2001; Biswas et. al., 2005;
Schwartz et al., in press). We describe the four most important principles.
Use Explicit and Well-Structured Visual Representations
A critical component of learning-by-teaching is knowing what one’s tutee has in
mind. Unlike human learners, we design the computer-based TAs to make their thinking
visible using well-structured visual representations and intuitive reasoning mechanisms.
For example, Betty’s representation of causal relations provides a well-structured graph
that is both intuitive and common among experts reasoning about causal propagation.
Betty’s visual representation helps students formalize and organize their own thinking.
For example, in one study, we asked college students to read a passage on metabolism.
Half of the students taught Betty and half wrote a summary of the passage. Students who
taught Betty adopted her knowledge representation, for example, by exhibiting more
complex chains of causal reasoning on a posttest.
Enable the Agent to Take Independent Actions
The second design principle is that TAs take independent actions based on what
they have been taught. It is not until students see Betty reason independently that they
can get useful feedback and models of reasoning. This opportunity to observe how one’s
student performs is much more informative than just teaching and not seeing the effects
of that teaching. In one study, for example, we found that students learned more when
they saw their agent perform compared to when the students taught their agent and then
solved the problems themselves instead of the agent. Moreover, it is quite motivating to
see how one’s tutee performs. For example, middle-school students had to help Betty
pass a test so she could join a science club. Students were eager to help Betty improve
and retake the test several times, something they are not always willing to do when they
need to retake a test themselves.

Model Productive Learner Behaviors
Another important component of learning-by-teaching is the need to monitor the
tutee’s performance, and to pay attention to new ideas the tutee might introduce. Young
children do not spontaneously know how to monitor thinking. Therefore, to help young
students learn these important skills, TAs model good learners (Biswas, Schwartz,
Leelawong, Vye, & TAG-V, 2005). Under specific conditions, Betty can spontaneously
offer a learning strategy or a concern. For example, as students build Betty’s map, she
occasionally starts to check her understanding. She then remarks (right or wrong) that the
answer she is deriving does not seem to make sense. These spontaneous prompts help
students reflect on what they are teaching, and hopefully like a good teacher, check on
their tutee’s learning progress.
Include Environments that Support Teaching Interactions
Teaching occurs in a larger context that includes goals such as preparing a student
for a test, a performance, a job, or a fun but complex game. The larger context of
teaching helps to motivate and focus a teacher. As we mentioned above, preparing Betty
to take a test is highly motivating and sets high standards. At the same time, a larger
context typically also provides resources that facilitate achieving these goals. There can
be books on the shelf, practice tests, model answers, a homework hotline, a nearby friend,
and so on. In the computerized context of TAs, we build environments that provide
students with a number of learning resources. Betty can appear in an environment that
includes searchable hypertext explanations of key topics. Betty can also appear in an
environment with a Mentor Agent who provides a combination of domain knowledge and
feedback, plus offers tips on how to monitor Betty’s knowledge. Importantly, the
learning-by-teaching environments should not undermine students’ need to actively
decide how to use the resources and feedback available in the environment. In one study,
we compared 5th-grade students who had to make decisions about how to use resources
and feedback versus students who were told exactly what to do. Students who had to
make decisions about resource use and teaching were able to learn more effectively a
month later on their own, even when they were no longer learning in the context of
teachable agents (Biswas, Schwartz, Leelawong, Vye, & TAG-V, 2005).
ADDITIONAL LEARNING-BY-TEACHING POSSIBILITIES
The four core principles are common to all the TAs. At the same time, we can
extend each TA in different directions to further support learning. TAs have a modular
software architecture that allows us to repurpose and combine them into different
applications. People learn by teaching in many ways, and in the following section, we
provide some examples of how we repurpose Betty to optimize different aspects of
learning-by-teaching

Enhancing Assessment through Learning-by-Teaching
The independent performance of TAs offers excellent assessment opportunities.
In the preceding examples, we described how students could assess their agents (and
themselves) one-on-one. We have extended this capability into a front-of-the-class
assessment system so students can see the performance of their own agent, as well as that
of their classmates. The agents created by each student can be displayed side-by-side.
The classroom teacher can ask a question of all the agents simultaneously. A hidden
expert map determines the correct answer. The results are tabulated and indicated by
color coding. The classroom teacher can zoom in on a map to see why an agent gave the
answer it did, and then compare it to another map as shown in Figure 2,. In a formal
study with college classes, we have found that the front-of-the-class system helps
students learn the structure of a domain better than just seeing the performance of their
own agent.
Figure 2. Front of class quiz system for showing agents perform. The classroom
teacher can ask all the agents the same question simultaneously. The top of the panel
indicates whether each agent answered the question correctly with red and green
highlighting. The bottom of the panel shows how the classroom teacher sets up a class
discussion by highlighting how two different agents reasoned about the same problem

Rethinking Homework
With TAs, we can reach beyond the classroom. The Triple-A Game Show is an
internet application designed to change homework practices (Figure 3). The student
teaches his or her agent and customizes the agent’s look. The student and agent then
participate in an on-line game show with other students and their agents. Students can log
on from home or school. The game host asks agents to answer questions and show their
thinking. Students can see their own agent reason, and also observe the reasoning of
other students’ agents. The application also includes a chat environment so students can
discuss and cheer (or jeer) an agent’s performance. Our hope is that students will find
this socially rich environment both engaging and educational, and it will prepare them for
their lessons in school the next day or week.
Figure 3. A customized agent performs in an on-line game show with other agents.
Videogames that Capture the Many Ways that People Learn
People learn in many ways. In our final example, we combine many forms of
learning and teaching interactions into a guided-discovery video game called Pumpkin
World (Blair & Schwartz, 2005). Figure 4 provides a sample screen shot. Betty takes the
form of an embodied agent, and students teach her how to grow giant pumpkins so she
can win contests at a pumpkin festival. In Pumpkin World, Betty can take actions
(instead of just answer questions). Betty, for example, reasons about what nutrients to add
to the soil to increase pumpkin growth. As she takes actions in the game world, the
student can see the functional consequences as the pumpkin grows or wilts. The

environment and narrative of the game support many other types of learning interactions
and resources as well. For example, students learn about the role of nitrogen through
experimentation; they learn about phosphorous by observing another agent; they learn
about “energy” by being told by another agent; and so on. Connecting TAs with video
games is not only motivating, it also provides a way to bring together a variety of
resources and interactions into one coherent problem solving environment.

CONCLUSIONS
Using the common wisdom that people learn by teaching we have developed
pedagogical agents that students must explicitly teach. Unlike most instructional
technology where the computer teaches and tests the students, with TAs students teach
and assess computer. Students readily adopt the fiction of teaching an agent. They find it
motivating, and it helps them to organize otherwise complex human-computer
interactions and learning tasks. Our research has also shown valuable learning benefits of
teachable agents. Perhaps, equally important, we have found that the TA metaphor to be a
generative source of novel ideas for creating new forms of instruction.
ACKNOWLEDGMENTS
This material is based upon work supported by the National Science Foundation under
Grants REC-0231946, REC-0231771, and Stanford’s Media-X. Any opinions, findings
and conclusions or recommendations expressed in this material are those of the authors
and do not necessarily reflect the views of the National Science Foundation or Media-X