Artificial intelligence (AI) tools like ChatGPT, despite their advanced capabilities and vast training datasets, still fall short when it comes to understanding and representing sensory and motor experiences as comprehensively as humans do, according to a recent study published in Nature Human Behaviour. The research, led by Qihui Xu, a postdoctoral researcher in psychology at The Ohio State University, highlights the limitations of large language models (LLMs) in grasping the full richness of human concepts, especially those rooted in sensory and physical interactions. The Study’s Context and Methodology Xu and her team conducted a comparative analysis of human and machine understanding of over 4,442 words, ranging from tangible objects like "flower" and "hoof" to abstract concepts like "humorous" and "swing." The study employed two measures to assess the alignment between human and LLM representations of these words: the Glasgow Norms and the Lancaster Norms. The Glasgow Norms evaluate words on nine dimensions, including emotional arousal, concreteness, and imageability. For instance, a flower might be rated highly on how emotionally arousing it is and how easily it can be visualized. The Lancaster Norms, on the other hand, focus on how concepts are linked to sensory and motor experiences, such as the sense of smell and bodily actions involved with the word. Key Findings When comparing human and LLM responses, the study found that AI models, including those from OpenAI (GPT-3.5 and GPT-4) and Google (PaLM and Gemini), performed well on words with no direct sensory or motor connections. However, AI struggled significantly with words that involve sensory experiences or physical actions. For example, while LLMs can approximate the concept of "smell" associated with a flower, they cannot fully grasp the intricate combination of sensory experiences—like the aroma, the tactile sensation of petal silk, and the emotional joy—that humans associate with a flower. This disparity arises because LLMs primarily rely on language and sometimes images, lacking the multidimensional sensory and motor experiences that humans bring to bear on their understanding. Implications for AI-Human Interaction The findings have broader implications for how AI and humans interact. If AI tools cannot fully comprehend human concepts, their ability to communicate effectively and empathetically with humans may be limited. As Xu points out, "If AI construes the world in a fundamentally different way from humans, it could affect how it interacts with us." Human understanding of concepts is deeply rooted in personal experiences and physical interactions, creating a rich, multi-faceted cognitive framework. LLMs, by contrast, develop their knowledge from vast amounts of textual data, which, while extensive, lacks the depth and nuance provided by real-world sensory and motor inputs. This discrepancy means that even after consuming significantly more text than a human would in a lifetime, AI remains unable to fully capture certain human concepts. Future Directions Despite these limitations, the study suggests that LLMs are improving. Models trained with both text and images, for instance, performed better in representing concepts related to vision compared to text-only models. This improvement indicates that integrating more diverse forms of input, such as sensor data and real-world interactions via robotics, could enhance AI's ability to understand and interact with the physical world in a more human-like manner. Industry and Expert Insights Industry insiders and experts are divided on the implications of this study. Some argue that while AI's current limitations in sensory and motor understanding are significant, ongoing advancements in multimodal learning and sensor integration could bridge this gap. Dr. Yingying Peng, a co-author from the Hong Kong Polytechnic University, emphasizes that "the potential for AI to gain a more holistic understanding of the world is vast, and we are only scratching the surface." Meanwhile, companies like OpenAI and Google continue to invest heavily in research aimed at making their AI models more sophisticated. OpenAI's development of GPT-4, which shows improvements in sensory-related tasks, and Google's exploration of multimodal models like PaLM and Gemini, highlight the industry's commitment to overcoming these challenges. Company Profiles The Ohio State University: Known for its interdisciplinary research and academic excellence, OSU has contributed significantly to fields like psychology and AI. Qihui Xu’s work exemplifies the university's focus on understanding the intersection of human cognition and machine learning. OpenAI: A leading AI research organization, OpenAI aims to develop and deploy AI in ways that benefit humanity. Their models, including GPT-3.5 and GPT-4, have set new standards in natural language processing, although this study indicates areas for further improvement. Google: A global technology leader, Google’s AI division is at the forefront of developing advanced AI systems. Models like PaLM and Gemini are part of Google’s efforts to create more versatile and context-aware AI tools. In conclusion, while AI has made remarkable strides in mimicking human language and understanding, it still faces significant challenges in fully representing sensory and motor experiences. However, the continuous evolution and integration of diverse data types offer a promising path toward more comprehensive AI models.