Intelligent Tutoring Systems

Anderson, Boyle, and Reiser's 1985 paper is one of the foundational works behind intelligent tutoring systems. Its central claim is striking even today: advances in cognitive psychology, artificial intelligence, and computer technology made it feasible to build computer tutors that approach the effectiveness of intelligent human tutoring in constrained domains. The paper describes computer tutors grounded in ACT theory, a cognitive architecture that models how learners acquire procedural knowledge. The authors discuss tutor systems for geometry proof and LISP programming, showing how AI could be built around explicit models of expert problem solving and student learning.

The importance of the paper lies in how it defines the architecture of tutoring as more than content delivery. A tutor needs a model of the domain, a model of the learner, and pedagogical rules for intervention. In geometry or programming, the system can compare a learner's actions against a model of correct strategies, infer what production rules the learner has or has not mastered, and provide feedback at moments when the learner is likely to benefit. This approach became central to later cognitive tutors, knowledge tracing, step-based feedback, and mastery learning systems.

For a modern Research News audience, this paper is useful because it reminds us that AIED did not begin with generative AI. Many of today's questions about AI tutors were already visible here: What knowledge should the system model? How should it diagnose learner understanding? When should it intervene? How can feedback be individualized without replacing the teacher? The paper's technical language is older, but the design questions remain current.

The historical contribution is also methodological. Anderson and colleagues show that tutoring systems can be designed by starting from a theory of cognition, decomposing expert performance, and building software that responds to learner actions in relation to that theory. This is different from adding generic automation to a lesson. It asks the designer to specify what counts as knowledge, what errors reveal, and what kind of feedback advances learning. That discipline is valuable when reviewing newer AI tutor proposals. If a modern system cannot say what learner state it is modelling, what misconception it is diagnosing, or why a hint is pedagogically appropriate, then its intelligence may be more conversational than instructional.

The paper also offers a useful contrast with today's large language models. LLMs generate flexible language but often lack explicit student models, domain models, or durable estimates of mastery. Classic intelligent tutoring systems had narrower domains but clearer pedagogical structure. AIEDHK can use this contrast to ask how current AI systems might combine the flexibility of language models with the instructional discipline of cognitive tutors. For multilingual classrooms, the legacy matters because it points toward curriculum-aligned AI that knows subject structure, misconceptions, and practice sequences. It also cautions against treating conversational fluency as proof of tutoring competence. The enduring lesson is architectural: effective tutoring requires explicit commitments about knowledge, evidence, timing, and instructional purpose.