Do students learn better with texts that are "slightly harder to read" (i.e., disfluent)? Previous research yielded discrepant findings concerning this question. To clarify these discrepancies, the present study aimed at identifying a boundary condition that determines when disfluent text is, and is not, beneficial to learning. This boundary condition is knowledge about whether a test will follow (high test expectancy) versus not (low test expectancy). Participants with high test expectancy may already engage in effortful processing, so that making text harder-to-read (disfluent) might not change their processing mode any further. Thus, particularly when no test is expected, disfluency is supposed to exert its beneficial effect. This assumption was tested in a 2 × 2 design (N = 97) with text legibility (fluent vs. disfluent) and test expectancy (low vs. high) as factors, and learning outcomes (retention, transfer) and learning times as main dependent variables. Results revealed that high test expectancy led to better learning outcomes (for retention and transfer), but disfluent text did not. Unlike expected, there was no interaction between the two factors. Moreover, both high test expectancy and disfluency led to longer learning times, resulting in a lower efficiency when learning with disfluent compared to fluent text. Hence, the present results further question the stability and generalizability of a positive disfluency effect on learning, because only high test expectancy--but not disfluency - stimulated better learning through more effortful processing the way it was supposed to.