ZotGPT to Help in Design of Nicotine Cessation Bot

In January 2024, UC Irvine launched ZotGPT Chat, a customized generative AI tool. Now, Ian Harris, a professor of computer science in the Donald Bren School of Information and Computer Sciences (ICS), is leveraging the new tool in a project that aims to help people address nicotine addiction.

The project is part of ongoing research into online interventions for smoking cessation led by Connie Pechmann, a professor of marketing in the Paul Merage School of Business. Harris is working to advance findings from Pechmann’s earlier study of a novel, Twitter-based social media intervention for smoking cessation called Tweet2Quit. The intervention targets online support groups for people who are trying to quit smoking.

“When people didn’t get responses on important tweets, they were more likely to quit the support group and go back to cigarettes,” explains Harris. “So Pechmann wants a chatbot that will respond when humans aren’t available.” That’s where Harris, with his expertise in natural language processing and machine learning, comes into play. In developing a chatbot with appropriate responses, Harris says ZotGPT could potentially help with several different tasks.

“The first thing we’re doing is classifying the tweets,” he says, clarifying that Twitter is now X but also that the platform itself doesn’t matter; they’re simply inputting text-based messages. “From Pechmann’s previous study, we have around 25 different types of tweets that we want to classify to help determine the right kind of response.” So Harris and computer science Ph.D. student Salar Hashemitaheri are using ZotGPT to help with this task, experimenting with how best to format the prompt. “You have to define the class of tweets, or give it examples of the class,” he says. “And then we ask, ‘Does this new tweet fit into the class?’”

Eventually, they hope to be able to submit thousands of messages for automatic classification and appropriate response generation. “The natural language processing field moves so darn fast that it has already changed since we started this project,” says Harris, who at some point would like to use ZotGPT to also provide the responses. However, that would require access to the Application Programming Interface, something he doesn’t currently have.

“A better way to go at it is to actually fine-tune the model — basically, to train ZotGPT itself to be better,” says Harris. “We can do that with ChatGPT, but it costs a lot of money. If we could do that with ZotGPT, we could make it generate the appropriate responses.”

For now, Harris and Hashemitaheri are focused on message classification. Once completed, undergraduate students will verify the classifications, and then Harris will work on how best to generate responses. The goal is to create a chatbot so that Pechmann can run another experiment, testing whether the chatbot helps people quit smoking. Harris will continue to explore what role ZotGPT can play in this cross-disciplinary collaboration with potentially life-saving health benefits.

— Shani Murray