In the context of escalating hate speech and polarization on social media, this study investigates the potential of employing responses generated by Large Language Models (LLM), complemented with pertinent verified knowledge links, to counteract such trends. Through extensive A/B testing involving the posting of 753 automatically generated responses, the goal was to minimize the propagation of hate speech directed at Ukrainian refugees in Poland.

The results indicate that deploying LLM-generated responses as replies to harmful tweets effectively diminishes user engagement, as measured by likes/impressions. When we respond to an original tweet, i.e., which is not a reply, we reduce the engagement of users by over 20% without increasing the number of impressions. On the other hand, our responses increase the ratio of the number of replies to a harmful tweet to impressions, especially if the harmful tweet is not original. Additionally, the study examines how generated responses influence the overall sentiment of tweets in the discussion, revealing that our intervention does not significantly alter the mean sentiment.

This paper suggests the implementation of an automatic moderation system to combat hate speech on social media and provides an in-depth analysis of the A/B experiment, covering methodology, data collection, and statistical outcomes. Ethical considerations and challenges are also discussed, offering guidance for the development of discourse moderation systems leveraging the capabilities of generative AI.

Joint work with: Jakub Podolak, Szymon Łukasik, Paweł Balawender, Jan Ossowski, and Katarzyna Bąkowicz.