When we asked GPT-3, a powerful artificial intelligence language system, whether a paper map or a stone would be better to fan life into coals for a barbecue, it chose the stone while people would typically choose the paper. Similarly, when asked to choose between a thermos and a hair pin to smooth out wrinkles, it suggested a hairpin while humans typically choose a thermos. GPT-3 even chose a hamburger bun to cover the hair in place of a paper sandwich wrapper. Why is it then that GPT-3 makes choices different than humans even though it is an advanced language model? This is because GPT-3 does not think and understand like humans do, and this is partly due to its lack of a physical body.
The meaning of words and sentences are closely correlated with human body and its different abilities such as action, emotions, and perception. In order to understand the world, it requires an “embodied” cognition, where people can understand objects by their physical properties such as foldable, scrunched, and covering. Take for example the paper sandwich wrapper and its countless uses, which are all created with the use of a human body and experience.
GPT-3, its successor GPT-4, Bard, Chinchilla, and LLaMA are all advanced language models developed to increase the understanding of words and sentences. However, without a physical body, these AI are unable to comprehend or determine the different properties, known as affordances, of an object like humans can. And even after adding a visual input feature to GPT-4, it is still far from understanding language like humans.
In order for artificial intelligence to ever understand language like humans, there needs to be a humanlike body, a set of senses, a purpose, and a way of life programmed into them. There are begun to be experiments with large language models that provides an access to physical environment and use simulations and robotic action plans, but this is only the first step towards an embodied AI.
ChatGPT, created by Open AI, is becoming a popular language model for many everyday scenarios. It is able to produce reasonable sentences, essays, and poems. Yet, despite its advances, it is important to note that ChatGPT does not understand the language it speaks, and it is not sentient.
Arthur Glenberg, a psychology researcher and Cameron Robert Jones, a Cognitive Science doctoral student, both have worked on understanding the discrepancies between artificial language models and the understanding of humans. Through their research, they found that AI still lacks the ability to comprehend the world in the same way humans do, and many of the everyday solutions we take for granted are not accessible to it.
In conclusion, language models like ChatGPT and GPT-3 are still extremely far from understanding language the same way humans do. They do not have an embodied body to understand the physical properties of objects. All they can do is to predict the next words based on statistical regularities of large data sets, which can give reasonable outputs but far from the same level of understanding and Sentience as humans.
