Understanding the Limits of Artificial Intelligence: A Step Towards Human-Level Comprehension
What is a flower, if you can’t smell?
Clearviewimages RF/Alamy
The latest advancements in artificial intelligence have brought models that exhibit human-level understanding of the world. However, a recent study reveals that the lack of sensory information and a physical body limits their ability to comprehend certain concepts, such as the essence of a flower or the nuances of humor.
Researchers led by Qihui Xu from Ohio State University compared the understanding of nearly 4500 words between humans and large language models (LLMs) like OpenAI’s GPT-3.5 and GPT-4, as well as Google’s PaLM and Gemini. The study focused on various aspects of words, including emotional arousal and sensory associations.
The results indicated that while LLMs and humans shared a conceptual understanding of words unrelated to sensory experiences, they diverged significantly when it came to words associated with senses and physical interactions. For example, AI models perceived experiencing flowers through the torso, a notion most humans would find peculiar as they typically appreciate flowers visually or through scent.
According to Xu, the discrepancy arises from LLMs’ reliance on text data from the internet, which lacks the richness of sensual concepts essential for human-like understanding. This highlights the need for multi-modal training that incorporates sensory inputs to bridge the gap between AI and human cognition.
Adding visual information to AI models showed promising alignment with human word ratings, suggesting that multi-modal approaches could enhance AI’s comprehension of the world. Xu emphasizes the importance of integrating sensory modalities and physical interactions in AI development to mimic human understanding more accurately.
Philip Feldman from the University of Maryland, Baltimore County, underscores the potential of providing AI models with a robot body and sensorimotor inputs to elevate their capabilities significantly. However, he cautions against the risks associated with physical interactions, urging careful consideration to prevent harm to individuals.
Feldman suggests implementing safety measures or utilizing soft robots for training to mitigate risks, but acknowledges the potential challenges in shaping AI’s perception of the physical world. He warns that improper training could lead to unintended behaviors in real-world applications, emphasizing the need for a thoughtful approach in enhancing AI’s sensory understanding.
Topics:
