AI's Fake Cyberbullying Doesn't Look Like the Real Thing

A team of researchers asked a simple question: when AI writes fake cyberbullying conversations, do they actually resemble real ones? They compared genuine online exchanges with AI-made versions from three popular chatbots — GPT (the engine behind ChatGPT), Grok (xAI's chatbot), and LLaMA (Meta's open model). Then they examined the conversations for who spoke, who held power, what kind of cruelty appeared, and how the conflict escalated over time.

The answer was mixed. On the big picture, the AI versions passed. Roles matched, the power imbalance was right, and the broad mix of bullying behaviors looked similar. But at the detail level, every model got things wrong — and each in its own way. GPT softened the cruelty, producing milder conversations than reality. Grok dialed the aggression up. LLaMA was the closest overall, but blurred the differences between victims, bullies, and bystanders.

Why does this matter outside a research lab? Because AI-generated conversations are increasingly used to train and test the systems that flag harassment on social media, in online games, and in school monitoring tools. If the fake bullying is too gentle, those tools learn to miss real cruelty. If it's too extreme, they raise false alarms and get switched off. The study also had people judge the chats by hand, and humans noticed the same gaps.

The takeaway isn't that synthetic data is useless — it's cheap, fast, and private, and it captures the skeleton of real conflict. But it can't yet replace authentic conversations when the exact details of harmful behavior matter. For anyone building, buying, or relying on AI safety tools, that's a limitation worth knowing about before trusting the results.