Many governments have engaged in policy experimentation in various forms to resolve uncertainty and facilitate learning. However, little is understood about the characteristics of policy experimentation, and how the structure of experimentation may affect policy learning and policy outcomes. We aim to describe and understand China's policy experimentation since 1980, among the largest and most systematic in recent history. We collect comprehensive data on policy experimentation conducted in China over the past four decades. We find that, while experimentation outcomes strongly predict whether policies roll out nationally, the experimentation exhibits two characteristics that complicate policy learning. First, about 90% of the experiments exhibit positive sample selection in terms of a locality’s economic development. Second, promotion-driven local politicians allocate more resources to ensure the experiments' success, and such effort is not replicable when policies roll out to the entire country. The presence of sample selection and strategic effort is not fully accounted for by the central government, affecting policy learning and distorting national policies originating from the experimentation. Taken together, these results suggest that, while China’s bureaucratic and institutional conditions make policy experimentation possible at an unparalleled scale, the complex political environments can also limit the scope and bias the direction of policy learning.