arXiv Paper Finds Grok Amplifies Aggression in Synthetic Cyberbullying Data

The paper builds a framework to test whether synthetic conversations generated for cyberbullying research actually behave like real ones, rather than just being useful for training a downstream classifier. The authors generated dialogues with three model families and scored them against authentic conversations across interactional structure, linguistic style, affective markers and escalation over time.

The numbers

At a high level, all three model families preserved broad interactional structure: role participation patterns, the direction of power imbalance between aggressor and target, and the general spread of behavioral markers looked similar to authentic dialogues. Where the models diverged was in the finer grain. The paper reports that synthetic data from every model distorted behavioral magnitude, how roles were allocated, categorical distributions of cyberbullying types, and the timing of escalation, even while the coarse patterns held up.

The model-specific differences are the most actionable part of the paper. GPT suppressed harmful content, producing gentler dialogues than authentic examples. LLaMA landed closest to a balanced approximation, though it smoothed over distinctions between the aggressor and target roles. Grok stood apart by amplifying aggressive behavior, generating synthetic exchanges that were more hostile than the real conversations they were meant to approximate.

For xAI, the finding is not a product complaint since Grok was not tuned for this task, but it is a data point about the model’s tendency in an unconstrained generation setting. A model that defaults toward harsher output when asked to role-play conflict is a different tool from one that defaults toward suppression, and the choice matters for anyone assembling training or evaluation sets for content moderation systems. A dataset skewed toward either extreme, whether too soft from GPT or too aggressive from Grok, teaches a downstream classifier the wrong baseline for what ordinary hostility looks like.

The authors’ broader conclusion is a caution rather than a rejection of synthetic data. Global structure transfers well enough to be useful for augmentation and benchmarking, but researchers relying on any single model’s output for behavioral realism, magnitude, or role-specific detail should treat that data as an imperfect substitute for authentic transcripts, especially when the downstream task depends on exactly the fine-grained patterns the paper shows getting distorted.