ChatGPT, an artificial intelligence (AI) language system, can spew out human-sounding sentences with great accuracy. But while ChatGPT can generate effective conversation-like messages, it is not able to understand language like humans do. A recent study highlighted this point as it found ChatGPT produced incorrect decisions on simple language scenarios. It confirmed that language understanding requires more than merely recognizing patterns in large sets of data.
AI engines, such as GPT-3 which powered the initial release of ChatGPT, can learn language from observing trillion instances of human language. But it is not enough for AI to know what comes after what. Human understanding of language carries the properties of being bodily and emotionally involved in the environment.
This was highlighted in the study which showed that AI does not consider the weight, feel, and use of objects when it comes to language comprehension. For example, it may not understand how to use a paper sandwich wrapper or a hamburger bun to cover one’s hair for a fast-food job. Humans, on the other hand, can make connections as they can access their embodied knowledge – their physical and emotional experience with the world.
The study team experimented with GPT-4, an AI system trained with images, to assess its language understanding. Interestingly, it was more successful than its predecessor GPT-3 in giving the right answers but it still performed significantly worse than humans.
The takeaway is that, for AI advantageously interpret language, it needs to have a body and be able to interact, perceive, and have emotions in the physical and social environments. Current AI language models focus only on learning from large data sets and, thus, do not possess the qualities of embodied cognition.
To address this lack, researchers are now attempting to integrate language learning with physical and interactive experiences like physics simulations and environment interaction. Such projects could pave the way towards progressing AI towards understanding language the way humans do.
ChatGPT has been a boon to many applications, but it is important to remember that, despite having human-like speech, it does not have the same level of comprehension as us.
GPT-3 and GPT-4 are large-language-model AI created by OpenAI, a company based in San Francisco, California. OpenAI is committed to creating super-intelligent AI that can benefit humanity. The company has developed a range of AI products and services, including language models and robotics, research software and cloud services. OpenAI has been successful and has continued to expand its efforts, raising hundreds of millions of dollars in investments.
Arthur Glenberg, Emeritus Professor of Psychology, Arizona State University, and a doctoral student in cognitive science were the researchers who undertook the study. Dr. Glenberg has spent over two decades studying how computers can process and understand language. His work enabled the creation of models that are used to this day. The doctoral student has been part of a team of researchers focusing on GPT-3. He also conducted experiments in search of AI that can understand language like humans do.
