Learning from experience, the cyclical interplay of thinking and doing, is increasingly important as organizations struggle to cope with rapidly changing environments and more complex and interdependent sets of knowledge. This paper confronts two central issues for organizational learning: (1) how is local learning (by individuals or small groups) integrated into collective learning by organizations? and (2) what are the differences between learning practices that focus on control, elimination of surprises, and single-loop incremental ?fixing? of problems with those that focus on deep or radical learning, double-loop challenging of assumptions, and discovery of new opportunities? We articulate these relationships through an analysis of particular learning practices in highhazard organizations, specifically, problem investigation teams that examine the most serious and troubling events and trends in nuclear power plants and chemical plants. We first distinguish a controlling orientation from a rethinking orientation, and illustrate learning practices with three case studies from the nuclear power and chemical industries and a questionnaire study of three nuclear power plants. We then extend our framework to create a four-stage model of organizational learning: (1) local learning by decentralized individuals and work groups, (2) constrained learning in a context of compliance with rules, (3) open learning prompted by acknowledgement of doubt and desire to learn, and (4) deep learning based on skillful inquiry and systemic mental models. These four stages contrast whether learning is primarily single-loop or double-loop, i.e., whether the organization can surface and challenge the assumptions and mental models underlying behavior, and whether learning is relatively improvised or structured. We conclude with a discussion of the stages, levels of learning (team, organizational, and individual), and the role of action, thinking, and emotion in organizational learning.