What ICAI makes possible, through its use of AI techniques, is an interactive learning session which gives great control to the learner as to how the session should unfold. Learner control of instruction, long-sought but not often realized in CAI, is a prime capability of ICAI, as has been indicated elsewhere (Duchastel, 1986). ICAI systems, however, are also meant gently to guide the student's exploration of the task domain being learned. To do this, they need to model explicitly both the student's ongoing status as it develops with respect to the domain as the learning session unfolds, and the very process of tutoring itself. Given the flexibility involved in ICAI, the student modelling process must induce a representation not only of what error the student may have just committed (as is the case in frame-oriented CAI) but of the student's global knowledge of the domain, including areas where knowledge is fuzzy, as well as outright misconceptions.