The Princeton University AI Lab team has introduced a new spatial intelligence benchmark called SSoN, specifically designed to test AI's ability to understand 3D space across static cross-view scenarios. This test strips away dynamic interaction and action-control interference, requiring models to identify, from four indoor images of the same scene, the anomalous image that does not belong to that scene. Results show that human accuracy reaches 83.3%, while mainstream vision-language models achieve at most 39.7%—with most falling below the four-choice random baseline—exposing core shortcomings in AI's spatial geometric reasoning: weak reasoning ability related to spatial transformations, invalid reasoning explanations for some correctly answered questions, and failure to establish a stable 3D spatial intuition. Subsequent intervention experiments show that textual prompt guidance or the introduction of traditional 3D reconstruction pipelines only marginally improves scores, whereas providing real camera parameters can substantially improve performance—though still not reaching a reliable correct-answer rate on more than half the questions; targeted supervised fine-tuning on this task can nearly double model accuracy on test questions in entirely new rooms. SSoN directly targets the core gap in current AI's cross-view 3D spatial integration, pointing the direction for spatial intelligence technology iteration.