ChatGPT and the Importance of Reasoning in AI
In the field of artificial intelligence (AI), ChatGPT has emerged as a leading language model renowned for its diverse capabilities. Whether it’s acing a bar exam, devising HR policies, or crafting movie scripts, ChatGPT seems to possess an edge. However, despite its impressive range of skills, there is a crucial aspect in which it falls short – the ability to reason like humans.
During a recent Q&A session, Dr. Vered Shwartz, an assistant professor in the UBC department of computer science, and Mehar Bhatia, a master’s student, highlighted the significance of reasoning in AI and the crucial role that diverse datasets play in training these models.
What does ‘reasoning’ mean in the context of AI?
Dr. Shwartz explained that large language models like ChatGPT learn by devouring millions of documents, essentially scouring the entire internet, to identify patterns and generate information. However, this means they can only provide information that has been explicitly documented. On the contrary, humans possess the exceptional ability to reason. We employ logic and common sense to derive meaning beyond what is explicitly stated.
Mehar Bhatia added that reasoning abilities are inherent to humans from birth. Through exposure to various situations, environments, and surroundings, we acquire knowledge that goes beyond formal teaching. In the near future, as AI models assume more tasks, it becomes impractical to manually program every single common sense rule into these models. Hence, the focus is on imbuing them with contextual understanding.
How do AI language models fall short?
Dr. Shwartz pointed out that while AI models do exhibit some form of common-sense reasoning, their capabilities are far from perfect. Relying solely on massive amounts of data for training will only take us so far. Human intervention is still necessary to train these models effectively, which includes providing the right data.
One area where AI language models face challenges is cultural bias. As English text on the web originates primarily from North America, English language models tend to exhibit a North American bias. This leaves them either unaware of concepts from other cultures or perpetuating stereotypes. In their research, Dr. Shwartz and Bhatia found that training a common-sense reasoning model with data from diverse cultures, such as India, Nigeria, and South Korea, resulted in more accurate and culturally informed responses.
Bhatia shared an example where the model, when exposed to culturally diverse training data, correctly suggested that a woman in Somalia receiving a henna tattoo might be preparing for her wedding. Previously, the model had erroneously suggested that she simply wanted to buy henna.
The significance of inclusive AI
Dr. Shwartz emphasized that language models are omnipresent. If these models are rooted in the values and norms of Western or North American culture, the information they provide regarding people from other cultures can be inaccurate and discriminatory. Additionally, people from diverse backgrounds who use products powered by English models might have to conform to North American norms, potentially limiting their overall experience.
Bhatia highlighted the importance of inclusive AI tools for everyone. With Canada being a culturally diverse country, the AI systems that shape our lives must not focus solely on one culture and its norms. Ongoing research aims to foster inclusivity, diversity, and cultural sensitivity in the development and deployment of AI technologies.
As AI language models like ChatGPT advance, it is crucial to bridge the gap in reasoning capabilities and address issues of cultural bias. By doing so, we can unlock the true potential of AI and ensure that these powerful tools benefit all of humanity, regardless of cultural background or perspective.
