Visual working memory for real-world objects is resistant to visual but vulnerable to semantic interference.

Visual working memory has greater capacity for maintaining meaningful objects than simple stimuli. This has been attributed to real-world object representations' higher dimensionality protecting their visual features from interference by those of other objects. However, object meaning could simultaneously create new, semantic sources of interference, but this possibility remains untested. Here, we addressed this question by quantifying low-level and high-level item similarity of real-world objects using a recurrent neural network model, as well as quantifying semantic similarity based on human descriptions of object features. Across four experiments, requiring participants to maintain the locations of six real-world objects in visual working memory, our results showed that high-level semantic but not low-level visual similarity between objects predicted memory errors. However, when object meaning was degraded, low-level similarity became a significant error predictor. These findings reveal that maintaining meaningful object representations in visual working memory grants protection from low-level feature interference but simultaneously increases susceptibility to high-level semantic interference. (PsycInfo Database Record (c) 2026 APA, all rights reserved).