Can Banning AI-generated Content Save User-Generated Q&A Platforms?

Abstract

The disruptions caused by large language models (LLMs), exemplified by ChatGPT, has prompted the implementation of moderating actions, including banning policies, across various sectors. Among these, the impact of such policies on user-generated content carries significant implications. User-generated content, particularly in Q&A formats, serves as a crucial learning resource for LLMs, making any changes to its quality and quantity influential in shaping the future learning trajectory of these models. This study investigates the effects of banning AI-generated content on online content platforms, using a natural experiment within the Stack Exchange network, where select forums enacted bans on AI-generated content. Through a generalized synthetic control approach, our analysis reveals adverse effects on platform vibrancy and usefulness, characterized by declines in both the volume and quality of questions and answers, as well as reduced instances of resolved questions. These outcomes can be attributed to a failure to consider two indirect effects in policy-making. First, the banning of AI-generated answers may lead to their partial replacement by human-generated answers of even lower average quality, as indicated by a comparison between changes in answers and the direct consequence of de-platforming answers with lower average quality. Second, although not the direct target of the ban, the act of raising questions, particularly high-quality inquiries, can be dampened by the reduced quantity and deteriorating quality of answers in a dynamic interplay, as evidenced by a causal mechanism analysis. These findings offer valuable insights and cautionary notes for future AI strategies on two-sided platforms and moderation efforts on content platforms.

Keywords: ChatGPT, Q&A platform, ban, AI-generated content, generalized synthetic control