Scaffolding is widely accepted as an effective strategy to use in learning environments. It is also agreed that for successful learning, scaffolding must fade to allow the learner to be in control of his/her own learning and acquire metacognitive strategies. Although scaffolding is used in almost all intelligent educational systems, fading is usually missing. In this paper we present an experiment whose goal is to study the appropriateness of fading. The study is performed in the context of the SQL-Tutor system, and the particular skill we focus on is the ability to select appropriate problems. We hypothesize that more able students would be better off when selecting problems on their own, and test whether this is valid. We also expect that for less able student, the most beneficial condition is faded problem selection: initially the system selects the problem for the student, giving explanations of why particular problems are good, and over time, the control over problem selection is given to the student. The results suggest that such an approach is effective.