Third generation computer tutors: learn from or ignore human tutors?

Current "second generation or "intelligent" computer tutors are approximately one-half as effective as human tutors. How will we develop the next generation of computer tutors that approaches human tutor effectiveness? Does success lie in understanding and emulating the performance of human tutors? If so, should we focus on natural language dialog or human tutor pedagogy? Alternatively, does computer technology afford effective instructional interventions, unavailable to human tutors? Can we modify learning activities and monitor student problem solving in ways that human tutors cannot.