E-commerce platforms increasingly deploy generative AI (GAI) to summarize large volumes of customer reviews. This study examines sentiment distribution bias in platform-provided GAI review summaries, defined as distortion in the relative representation of positive, neutral and negative sentiment compared with their proportional presence in the associated review corpus.
Grounded in information theory, we characterize four sentiment distribution bias patterns and develop an input-output framework linking source review characteristics and summary design features to sentiment bias. We analyzed data from an online travel platform covering 847 Tokyo hotels with GAI review summaries linked to 111,115 customer reviews to assess sentiment bias and conducted a complementary experiment to examine downstream consumer responses.
The platform's GAI review summaries exhibited predominant positive bias: positive sentiment was overrepresented, whereas neutral and negative sentiments were underrepresented relative to the review corpora. Lower source review volume, higher rating entropy and lower positive imbalance were associated with greater sentiment bias, whereas longer summaries, greater attribute granularity and lower attribute sentiment diversity were associated with lower bias. An experimental study showed that, relative to a low-bias summary, a GAI review summary with greater positive bias improved consumers' attitude toward the hotel and increased booking intention, despite being perceived as less helpful.
This research documents sentiment distribution bias in platform-provided GAI review summaries, identifies input and output conditions under which such bias varies and examines its consequences for consumer responses. The findings extend the online review and GAI literatures while informing platform governance and responsible GAI deployment.
