Chatbots such as Bard and ChatGPT are the new toys on the block that seem to be hogging the attentional bandwidth of people of all ages. From helping students complete assignments to answering questions about the Bhagavad Gita these chat tools provide responses that suggest human-like understanding of the nuances of language and culture. The large language model with its billions of parameters and its infinitely large datasets is being sold as the poster child of true lab grown intelligence. Are language models truly intelligent? Before one can form an opinion on this and other offerings of the AI effort, it is necessary to understand how these systems work.
Imagine you have been invited to the mythical country of Demolea where the language of official discourse is Guajalanga. You have less than one day to prepare for a series of high level discussions on a number of sensitive subjects, all of which are to be conducted in Guajalanga. You have never in your life spoken Guajalanga and you have no idea what its grammar is like, much less what its cultural connotations are. What would you do? You would probably do what anyone faced with a sudden foreign language test would — memorize as many questions and answers as possible, listen for keywords and expressions that match the questions and topics in your little guidebook and then reproduce the material to the best of your ability. While this approach might get you through those state meetings, you would probably not claim to be a speaker of Guajalanga or indeed that you have any understanding of the contents of your guidebook. And yet this is how AI language models are trained. Deviations from the example above are of quantity and speed.
Instead of a small guidebook you have the ever growing corpus of human language content on the internet, a correspondingly humungous memory required to store all possible combinations of words and phrases and a probability attached to the occurrence of these elements within the corpus (not to mention the massive underpaid human workforce used to train language models).
AI language models which have no understanding of the words they produce, follow a kind of copy-paste approach while responding to questions. The result is that these chatbots can end up producing gibberish such as links to websites that do not exist, descriptions of fake scientific concepts, threats, insults and gaslighting. However the key takeaway here is that a chatbot trained on the body of online language content is going to be limited to a large extent by the quality of this content. In this sense these language models mirror the trends and biases of our society, and could be said to carry out “human-like” discourse, biased or otherwise. While the danger of chatbots being used to generate fake news articles or to propagate biases is a real one, understanding how language models work could help motivate a more thoughtful and responsible conversational style among humans. AI models like children repeat what they hear and ought therefore to be consciously parented. There is a real opportunity here for AI, even its present rudimentary form, to induce positive transformation of human society and for human users to be active participants in the development of AI tools.
References
https://www.wired.com/story/bing-chatbot-errors-ai-future/
%2C%20low%2Dpaid%20human%20labor
https://www.vice.com/en/article/k7bmmx/bing-ai-chatbot-meltdown-sentience?
mc_cid=5a2bb2ac96&mc_eid=abdcc19d97
https://medium.com/madebymckinney/the-gender-bias-inside-gpt-3-748404a3a96c
https://timesofindia.indiatimes.com/life-style/spotlight/i-have-been-a-good-bing/articleshow/
98295541.cms?from=mdr
Views expressed by the author are personal and need not reflect or represent the views of the Centre for Public Policy Research.
Dr Monika Krishan's academic background includes a Master’s in Electrical Engineering from the Indian Institute of Science, Bangalore, India and a Ph.D. in Cognitive Psychology from Rutgers University, New Jersey, USA. Her research interests include image processing, psychovisual perception of textures, perception of animacy, goal based inference, perception of uncertainty and invariance detection in visual and non-visual domains. Areas of study also include the impact of artificial intelligence devices on human cognition from the developmental stages of the human brain, through adulthood, all the way through the aging process, and the resulting impact on the socio-cognitive health of society. She has worked on several projects on the cognitive aspects of the use and misuse of technology in social and antisocial contexts at SERC, IISc as well as the development of interactive graphics for Magnetic Resonance Imaging systems at Siemens. She is a member of Ohio University’s Consortium for the Advancement of Cognitive Science. She has offered services at economically challenged schools and hospitals for a number of years and continues to be an active community volunteer in the field of education and mental health