Missing Pieces: How Humans and GPT-4.1 Detect Absence and Predict the Unseen

We investigate the semantic reasoning capabilities of GPT-4.1 through a comparative study with human participants across three visual domains: object absence recognition, social inference, and temporal prediction. Our evaluation introduces a benchmark designed to assess the detection of absences and the prediction of unseen events. GPT-4.1 exceeds human performance in social context cases, where structured patterns support inference. In object-based cases, model performance is more variable, with failures occurring when crucial visual cues are subtle or embedded in dense scenes. Temporal reasoning is most challenging for both humans and GPT-4.1, though the model struggles when inference depends on narratively significant but less-prominent physical details.