OA

Positivity Bias in AI-Generated Summaries of User-Generated Content: Exploring Its Sources and Impact on Public Sentiment

Recently, platforms have been increasingly deploying generative AI (GenAI) to summarize user-generated content (UGC) into AI-generated summaries (AIGS). However, the potential bias in AIGS and its impact on the public remain inadequately examined. We used Weibo, a leading social media in China, as a case to investigate these important questions, focusing on public sentiments. Specifically, we explored whether AIGS are biased in representing emotions in UGC and whether such representation influences subsequent public sentiment. We empirically identify two sources of bias in the algorithmic processes underlying the production of AIGS from UGC: the sampling process, in which GenAI selects a subset of UGC, and the summarizing process, in which the summary is generated from the sampled content. Comparing emotions in AIGS, sampled UGC, and all UGC, we found evidence of bias in both processes. In our case, GenAI tends to favor positive UGC during the sampling process and produces summaries that further amplify this positivity, leading to an over-representation of positive sentiments in AIGS. Additionally, we utilized a Difference-in-Differences (DiD) design to explore AIGS in public sentiment dynamics. Findings suggest that AIGS alone are insufficient to influence public sentiment significantly. Overall, this study provides important implications for deploying GenAI in public online discussions.