Abstract
Large language models (LLMs) such as GPT-3 and their derivative products such as ChatGPT have garnered significant attention for their remarkable ability to process texts and conduct human-like conversations. Guided by the Diffusion of Innovation theory, this study examines the early discussions about LLMs on Twitter, specifically about ChatGPT and GPT-3, during the first three months following the release of ChatGPT. By utilizing topic structural modeling and sentiment analysis on a sample of 42,273 #ChatGPT tweets and 17,639 #GPT3 tweets, we explore how laypeople and technical professionals differ in their attitudes in the early stage of the adoption of LLMs. Our findings suggest that the discussion surrounding ChatGPT and GPT-3 primarily revolves around relative advantage and compatibility, with the majority of #ChatGPT conversations demonstrating negative sentiment and #GPT3 discussions containing more positive topics. The Twitter discussion using #ChatGPT is highly business-oriented, while the discussion of #GPT3 covers a broader range of topics in terms of the characteristics, applications, and potential ethical concerns of LLMs. This study offers implications for government agencies and policymakers, suggesting that further research is needed to fully understand the potential applications and risks of LLMs.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 While acknowledging the temporal limitations of our data collection, it is worth noting that the release of GPT-4, an advanced iteration of GPT-3, occurred after the completion of our data collection process.
2 Thematic frames refer to the underlying themes or concepts that are present in the text such as AI, conversational agents, and machine learning, while episodic frames refer to the specific events or situations described in the text including the release of ChatGPT and the impact of language models on education.
3 The top words are usually identified based on four indicators: Highest Prob, which shows the most probable words; FREX, which evaluates a word’s significance in the topic compared to its frequency in the corpus; Lift, which measures a word’s association with the topic; and Score, which provides an overall importance score for a word in a topic.
4 Each topic is represented by a horizontal line with a dot in the center, showing the level of uncertainty and coefficient of the hashtag on the topic. If the line does not intersect the central axis (0), it indicates a significant difference in the topic between #ChatGPT and #GPT3. Otherwise, it suggests no difference in the topic between these two hashtags. This rule applies to all figures in the current manuscript.
Additional information
Funding
Notes on contributors
Wenxue Zou
Wenxue Zou is an Assistant Professor in the Department of Communication, Media, & Culture at Coastal Carolina University. She got her doctoral degree in the Department of Communication and Journalism at Texas A&M University. She studies misinformation, culture, and social media from various approaches.
Jinxu Li
Jinxu Li is a Ph.D. student in the Department of Communication and Journalism at Texas A&M University. Her research focuses on health communication, social media, and artificial intelligence.
Yunkang Yang
Yunkang Yang is an assistant professor of communication at the Department of Communication and Journalism at Texas A&M University where he is also affiliated with the Data Justice Lab. His research interests include US right-wing media, disinformation, social media, and artificial intelligence.
Lu Tang
Lu Tang is a Professor at the Department of Communication and Journalism and Director of the Data Justice Lab, Texas A&M University. Her current research examines the use of emerging technologies in health promotion with a special emphasis on social justice and ethics.