In the realm of trying to decipher what makes humans uniquely perceptive, a new study hints at the limits of artificial intelligence (AI) when grasping human concepts like "flower." Researchers from The Ohio State University have unpacked, how AI, despite its sophisticated algorithms, still falls short of truly understanding sensory experiences and motor information associated with certain words. Their findings, showcased in a recent publication in Nature Human Behaviour, suggest that AI's perception of the world is fundamentally different from ours, potentially impacting AI-human interactions.
Qihui Xu, leading the study as a postdoctoral researcher in psychology at The Ohio State University, said "A large language model can’t smell a rose, touch the petals of a daisy or walk through a field of wildflowers," pointing to the inherent limitations machines confront without sensory and motor experiences. The research juxtaposed humans and large language models (LLMs) like GPT-3.5 and GPT-4 from OpenAI, and PaLM and Gemini from Google, in their conceptual understanding of over four thousand words spanning from the concrete to the abstract. While the researchers found that AI is on par with humans for words unlinked to the senses or physical actions, the distinction became clear when the concepts tied to our physical experiences, involving sight, taste, or touch, were at play.
The study employed measures known as the Glasgow Norms and Lancaster Norms to assess AI and human alignment regarding concepts and their sensory and motor relations. The Glasgow Norms explores ratings on dimensions such as how emotional or visual a word is. In contrast, the Lancaster Norms delve into how word concepts are related to sensory input and motor functions, for example, their connection to scent or actions from the torso. In both methodologies, the gap between human and AI word representation highlights the disparity in understanding visceral experiences, such as a flower's aroma or texture.
"From the intense aroma of a flower, the vivid silky touch when we caress petals, to the profound joy evoked, human representation of 'flower' binds these diverse experiences and interactions into a coherent category," Xu and her colleagues underscored in their research. AI, by contrast, operates primarily on language, and "language by itself can’t fully recover conceptual representation in all its richness," Xu noted via Ohio State News, marking a boundary for our digital counterparts.
While the findings may chart out the shortcomings of AI, they also offer a deeper insight into the unique capacities of human cognition and the intrinsic value that our physical interactions with the world contribute to understanding. The study, published today, prompts a re-evaluation of AI's role and its interface with human experiences, providing a conceptual framework that underscores the difference between humans and AI – we live and breathe our world in a way technology, as of yet, cannot replicate.
.jpg){kind=link}