You Are Allowed to Say More: ChatGPT Censorship on Controversial Topics and Contextual Prompting

Diego Jose Marquez, Kyrie Zhixuan Zhou, Yunpeng Xiao, Madelyn Rose Sanfilippo

Abstract:

With large language models (LLM) increasingly in the spotlight, their approach to censorship on topics like
immigration and conflict deserves a closer look. Our research investigates the role of censorship in LLMs, and how
these models manage controversial topics. We compare how acontextual and contextually prompted inputs shape
ChatGPT’s responses on subjects surrounding immigration policies and international conflicts, identifying context
as a critical factor in moderation behavior. While existing literature highlights LLMs’ ability to maintain fairness,
there is a gap in understanding how contextual prompting influences model responses and potential censorship
mechanisms. With systematic and contextual prompting, we reveal that contextually prompted models often deliver
more nuanced responses, potentially bypassing stricter moderation due to their evaluative nature. This study
contributes to the ongoing discourse on AI ethics by offering insights into improving LLM design to balance
objectivity and usability, ultimately informing policy guidelines for deploying AI in sensitive domains.