The Ohio State University researchers investigate the limitations of large language models, such as ChatGPT and Gemini, in understanding human concepts. Their study reveals that, while these models can excel in verbal tasks and pattern recognition, they fundamentally lack the sensorimotor grounding that enriches human understanding. Tested against over 4,400 words across various dimensions, including those based on sensory experiences, AI models fell short in grasping concepts like 'flower' or 'swing', underscoring the inadequacy of language alone for complete comprehension.
The study indicates that advanced AI models, although proficient in language tasks, cannot completely represent human concepts that rely on bodily experiences and sensory inputs.
Researchers found that despite AI's ability to identify linguistic patterns, concepts rooted in sensory interaction, like 'flower' or 'swing', are beyond their full understanding.
Collection
[
|
...
]