AI can be cyberbullying perpetrators: Investigating individuals’ perceptions and attitudes towards AI-generated cyberbullying

Cyberbullying is a critical social problem that can cause significant psychological harm, particularly to vulnerable individuals. While Artificial Intelligence (AI) is increasingly leveraged to combat cyberbullying, its misuse to generate harmful content raises new concerns. This study examines human perception of AI-generated cyberbullying messages and their potential psychological impact. Using large language models (LLMs), we generated cyberbullying messages across three categories (sexism, racism, and abuse) and conducted a user study (n = 363), where participants engaged with hypothetical social media scenarios. Findings reveal that AI-generated messages can be just as or even more harmful than human-written ones in terms of participants’ comfort levels, perceived harm, and severity. Additionally, AI-generated messages were almost indistinguishable from human-written ones, with many participants misidentifying AI-generated messages as human-written. Furthermore, participants with prior experience using AI tools consistently demonstrated higher accuracy in identification, while their attitudes towards online harm significantly influenced their comfort levels. This study emphasizes the urgent need for robust mitigation strategies to counter AI-generated harmful content, ensuring that AI technologies are deployed responsibly and do not exacerbate online harm.