LLM-generated cyberbullying data distorts social dynamics, study finds

A new study published on arXiv investigates the realism of synthetic cyberbullying data generated by large language models (LLMs) compared to authentic dialogues. Researchers found that while models like GPT, Grok, and LLaMA can replicate high-level interactional structures such as turn-taking and power dynamics, they consistently distort finer-grained social phenomena. The study highlights model-dependent biases, with GPT suppressing harmful content, Grok amplifying aggression, and LLaMA offering a more balanced but less distinct approximation of roles. The findings suggest that synthetic data is useful for understanding broad interactional patterns but is not a perfect substitute for authentic conversations when behavioral realism is critical. AI

IMPACT Highlights limitations of synthetic data for studying nuanced social behaviors in LLM-generated dialogues.

RANK_REASON Research paper analyzing LLM-generated data realism. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• generative pre-trained transformer
• Gotit.pub
• Grok
• Hugging Face
• LLaMA
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →