Some special problems associated with evaluating intelligent computer-assisted instruction (ICAI) programs are addressed. This paper intends to describe alternative approaches to the assessment and improvement of such applications and to provide examples of efforts undertaken and shortfalls. Issues discussed stem chiefly from the technical demands of the artificial intelligence field, which have tended to limit most evaluation efforts to first-party evaluation by project staff. ICAI evaluation should make use of a range of formative (e.g., componential analysis) and summative (e.g., cost analysis) evaluation methods with multiple criterion measures. Standardized tests have not proved sensitive enough in this area; domain-referenced tests are especially well suited to ICAI, because their success depends on experts' care in constructing detailed specifications of the knowledge domain. Individual differences in students' intelligence, cognitive styles, and state anxiety should also be considered. As evaluators sharpen their goals, they will be able to select the most relevant data to collect, and present useful instructional options. Special recommendations for ICAI evaluation include: (1) developing an expectation of evaluation; (2) rewarding evaluation participation; (3) increasing credibility of the evaluating team by encouraging expert participation; (4) adapting evaluation to specific features of ICAI development; (5) performing componential analysis of software under development; and (6) maintaining both a responsible and responsive approach. (LPG)