A regulator anticipates learning about the relation between environmental stocks and economic damages. For a model with linear-quadratic abatement costs and environmental damages, and a general learning process, we show analytically that anticipated learning decreases the optimal level of abatement at a given information set. If learning causes the regulator to eventually decide that damages are higher than previously thought, learning eventually increases abatement. Learning also favors the use of taxes rather than quotas. Using a model that is calibrated to describe the problem of global warming, we show numerically that anticipated learning causes a significant reduction in first period abatement and a small increase in the preference for taxes rather than quotas. Even if the regulator's initial priors about environmental damages are much too optimistic, he is able to learn quickly enough to keep the expected stock trajectory near the optimal trajectory. (c) 2005 Elsevier Inc. All rights reserved.