Why AI Language Models Like ChatGPT Struggle to Grasp the Essence of Flowers

Researchers at The Ohio State University have explored whether large language models (LLMs) like ChatGPT and Gemini can genuinely understand human concepts without sensory experiences. Their study reveals that while these AI models excel in identifying patterns and relationships in language, they fall short in grasping concepts that involve touch, smell, or physical interaction. Testing four major AI models on over 4,400 words, the researchers found strong alignment for abstract qualities, but significant limitations in sensory and motor-related concepts. For example, LLMs can recognize a flower visually but struggle to capture the full richness of human experiences tied to it. The study suggests that human understanding is deeply rooted in embodied experiences, which LLMs, trained primarily on text, cannot fully replicate. Future improvements may arise from combining AI with multimodal training, including visual and sensor data, but the unique richness of human sensory experience remains unmatched by current AI capabilities.

Source link