Theories like predictive coding propose that lower-order brain areas compare their inputs to predictions derived from higher-order representations and signal their deviation as a prediction error. Here, we investigate whether the macaque face-processing system, a three-level hierarchy in the ventral stream, employs such a coding strategy. We show that after statistical learning of specific face sequences, the lower-level face area ML computes the deviation of actual from predicted stimuli. But these signals do not reflect the tuning characteristic of ML. Rather, they exhibit identity specificity and view invariance, the tuning properties of higher-level face areas AL and AM. Thus, learning appears to endow lower-level areas with the capability to test predictions at a higher level of abstraction than what is afforded by the feedforward sweep. These results provide evidence for computational architectures like predictive coding and suggest a new quality of functional organization of information-processing hierarchies beyond pure feedforward schemes.