The rapid rise of generative AI language models, colloquially known as chatbots, has heralded a transformation of verbal communication – a realm that used to be the exclusive domain of humans. Much has been said about the capabilities of programs like ChatGPT: where they impress in some situations and fall short in others. Many say that they do not want to ‘talk to a robot’ because they find the conversation unengaging and that the chatbot lacks “personality”.
According to Sarah Rose Siskind, chatbot writer and founder of science comedy consultancy Hello SciCom, an AI's capability to converse is determined not only by the quantity but also by the quality of the data it was trained on. Many current-generation chatbots use the generative pre-trained transformer (GPT) language model, which rely on large amounts of data to train it to predict the most probable sequence of words and simulate human-created text.
“It's important that an AI program is able to capture human dialogue and is actually worth talking to,” Siskind says. “Because, at the end of the day, it doesn't matter how perfectly an AI can approximate human linguistic capabilities, if it's boring – nobody will want to talk with it.”
When it comes to training AI, both quantity and quality of information is important. A language model uses data from the conversations it has to increase its predictive capabilities. Siskind says that if people are having dry, unengaged conversations with AI, the data it will be trained on will have no familiarity, rapport, or entertainment value, making people less likely to talk with it.
“AI is only as good as the data it's trained on,” Siskind says. “So, to have better data, we need to have better conversations with the AI, which will lead to longer conversations. Quality leads to quantity. That's what we're trying to help accomplish.”
According to Siskind, properly trained artificially intelligent chatbots will not only be able to make a grammatically correct sentence, but will also be able to answer a question based on related text mostly found online. In order to provide better quality data to these generative AI models, Siskind says that humor is a really helpful inroad, as AI involves understanding predictability or expectations, while humor is all about upending expectations.
“Humor is like the antithesis of AI. Humor is about something that's unexpected, which often provides nuance. We humans like it because it surprises us and amuses us. The unexpected appeals to us because it's like play, which is part of social bonding, and this is a part of communication that AI currently does not have. Providing humor into these interactions allows for a neglected dimension of communication in the field of language models. However, there is one type of humor that is actually all about fulfilling expectations: impressions. If we think about it, AI is inherently funny, as it's a machine doing an impression of a human being, and it's getting pretty good at it.”
Part of Siskind's work at Hello SciCom involves creating training data and parameters for AI chatbots to make them more engaging and capable of holding better conversations. Drawing on their experience writing comedy for TV and public figures' speeches, Siskind and her team can help inject humor and personality into AI-powered conversational agents, making them more engaging to use. The team writes personality bibles, parameter conditions, consults on existing characters, and writes material to be used to train language models, giving them specific personalities that fit their purpose.
Siskind has gathered diverse writers from the Onion, Botnik, the Harvard Lampoon, TV, and various performance backgrounds. Together, they’ve written for chatbots like Sophia the celebrity robot at Hanson Robotics, published papers on rapport and robotics with Honda Research Institute of Japan, and helped create a generative AI-powered avatar sold at Sotheby’s for a half million dollars.
Aside from AI, Siskind and the Hello SciCom team also work with public figures and scientists to incorporate humor into their speeches and presentations, which makes science and tech more accessible and interesting to laypeople. Studies have shown that humor increases retention of information, and this is especially important today with the growing level of scientific illiteracy of the American public.
Siskind says, “Science is responsible for just about everything we know about the world around us. Science is the ‘why’ and tech is the ‘how.’ But people aren't going to remember the information experts give them if they're bored or asleep.” In this climate of distrust, Siskind believes lightness is the best disinfectant.
Media contact:
Name: Sarah Rose Siskind
Email: [email protected]
Original Source of the original story >> Why AI Shouldn’t Be Boring
