The rapid ascension of large language models and their multimodal variants has led to widespread claims of near-human or even superhuman intelligence, particularly in areas requiring extensive knowledge and complex reasoning. However, a new, highly focused study introduces a significant counter-narrative, exposing a fundamental and surprising frailty in today’s most powerful AI systems: they struggle profoundly with basic visual tasks that human toddlers master before they learn to speak. The research, which utilized a novel benchmark called "BabyVision," found that even the most advanced multimodal models, despite scoring high on expert-level knowledge tests, are routinely outperformed by the average three-year-old child on simple problems of perception and spatial awareness. This stark performance gap points to a critical flaw in current AI architecture and methodology, suggesting that the path to true Artificial General Intelligence (AGI) is obstructed by a lack of basic visual grounding.[1]

The "BabyVision" benchmark, developed by researchers from institutions including UniPat AI and Peking University, was constructed to test four core categories of visual primitives that developmental psychology research indicates humans acquire early in life: fine-grained visual discrimination, visual tracking, spatial perception, and visual pattern recognition. The total benchmark comprises 388 unique tasks designed to be independent of linguistic priors, which are the massive text-based datasets that give most current AI models their conversational brilliance. For comparison, the AI models were tested against a cohort of 80 children across different age groups, alongside human adults. The human adult baseline performance across all categories approached a near-perfect 94.1 percent accuracy.[1]

The results reveal a substantial chasm between human and machine visual comprehension. The best-performing proprietary model tested, Gemini-3-Pro-Preview, achieved an overall accuracy of just 49.7 percent. While this score was sufficient to beat the average three-year-old, it still lagged noticeably behind the performance of typical six-year-olds by approximately 20 percentage points. Other frontier models fared significantly worse. GPT-5.2 managed 34.4 percent, and Claude 4.5 Opus scored a mere 14.2 percent. Open-source models struggled even more, with the top performer in that group, Qwen3VL-235B-Thinking, reaching only 22.2 percent. A significant finding was that most frontier AI models tested scored below the average for three-year-olds, confirming that the vast majority of today's cutting-edge systems lack the visual common sense of a preschooler.[1]

The failures were particularly stark in tasks requiring geometric precision and continuous spatial tracking. For example, on a specific task involving counting 3D stacked blocks, the best AI model achieved an accuracy of only 20.5 percent, while human participants scored 100 percent. Similarly, in the "Lines Observation" task, which requires tracing lines through a complex pattern of intersections, most models scored zero percent accuracy. This task, which is trivially solved by a child, highlights the models' inability to maintain spatial coherence and follow a logical path continuously. The models demonstrated significant dips in performance for visual tracking, such as mazes, and spatial perception, like mentally unfolding a cube, indicating these are major stumbling blocks that remain unsolved by current architectures.[1][2]

The researchers attribute the widespread failures to a systemic problem they term the "verbalization bottleneck." Modern multimodal models typically operate by translating visual input—the image—into a language-based representation before they perform any reasoning or generate an answer. This process, which privileges the textual modality, causes critical geometric and non-verbal spatial information to be lost during the translation, effectively creating a "visual blind spot" for the AI. The models are, in essence, trying to describe a complex visual puzzle with language alone, leading to over-verbalization of geometry and a failure to grasp the exact contours or fine-grained differences. This reliance on language-based shortcuts, rather than robust visual understanding, is a crucial limitation.[1][2]

This research reinforces findings from other recent studies that have also pitted advanced AI against human intuition. For instance, a separate study evaluating vision-language models on classic visual puzzles known as Bongard problems found that top-tier models like GPT-4o achieved an accuracy of only about 17 percent, while human participants neared 84 percent. The cumulative evidence suggests that the ability to perceive, reason about, and innovate with physical and geometric concepts—skills developed through embodied, sensorimotor experience in humans—is absent in current data-driven AI. Toddlers naturally learn compositionality by integrating vision, action, and language through real-world, embodied interactions, a learning mechanism that appears to be key to visual grounding.[3][4]

For the AI industry, these results carry profound implications. The development of trustworthy AI for high-stakes applications like autonomous vehicles, medical imaging, and robotics depends entirely on reliable visual reasoning and common sense. A system that cannot accurately trace a line through an intersection, count objects in a specific spatial configuration, or consistently identify subtle visual differences is fundamentally unreliable in complex, real-world environments. The study suggests that to progress toward AGI, the industry must pivot away from a language-first processing paradigm. Future research and development must focus on an "early fusion" approach, where visual and language training are integrated from the very beginning, or perhaps adopt architectures that more closely mimic the human brain’s distinct, yet interconnected, systems for vision and language. Until AI can master the visual acuity and spatial reasoning of a three-year-old, its "intelligence" will remain a narrow, language-dominant parlor trick with a glaring weakness in the face of physical reality.[1][2][5]