The aim of this paper is to examine the learning processes that subjects undertake when they start using gaze as computer input. A 7-day experiment with eight Japanese students was carried out to record novice users' eye movement data during typing of 110 sentences. The experiment revealed that inefficient eye movements was dramatically reduced after only 15-25 sentences of typing, equal to approximately 3-4 h of practice. The performance data fits a general learning model based on the power law of practice. The learning model can be used to estimate further improvements in gaze typing performance. Our experimental results encourage the use of gaze-based interfaces for severely disabled people. This paper provides a taxonomy for gaze actions on dwell time-activated keys and it presents a method by which the learnability of gaze interfaces may be documented. (Contains 8 tables and 5 figures.)