Human Gaze Reveals a Task-Dependent Link Between Vision Model Depth and Human Visual Cortex

Shaped by survival pressures, human visual attention allocates limited processing resources to behaviorally relevant information, offering a guiding principle for adaptive artificial vision. Do artificial vision systems encode human-like visual saliency patterns? It remains unclear where fixation-relevant information becomes accessible within model hierarchies and how this localization corresponds to human visual cortex. We introduce a unified framework measuring fixation accessibility and cortical predictivity through separate linear readouts of the same frozen representations. We evaluate four prespecified depths in fifteen diverse vision models, comparing free-viewing and target-search fixations on matched images and using independent fMRI data as a fixed cortical reference. Within models, fixation accessibility and cortical predictivity covary across depth, with the strongest correspondence in intermediate visual, ventral, and lateral brain regions. Search shifts relative fixation accessibility toward later representations: cortical correspondence decreases for V1–V3, remains high for hV4 and midventral cortex, and increases for higher ventral, lateral, parietal, and category-selective regions. These findings establish representational depth as a task-dependent link between fixation accessibility and cortical predictivity, providing a functional criterion for interpreting human-like vision models. We also identify candidate cortical sources of gaze-relevant information and propose that goals guide selection partly by changing which representational levels supply useful information. This provides a testable design principle for artificial systems that recruit feature depths as goals demand.