As we continue our ChatGPT experiments, in one experiment today we were using ChatGPT to summarize a Russian news broadcast when we encountered the following warning that "This content may violate our content policy." The interaction and influence of the conversational health and moderation layers on public-access LLMs is poorly understood when it comes to using such models to understand global news coverage from conflict environments, repressive regimes and cultures whose values differ from traditional "Western" norms. As some of our examples have demonstrated, models may even refuse to answer questions about such content, posing unique challenges to codifying and understanding those narratives for fact checking, countering, narrative modeling, etc. As the example below shows, even the simple task of summarizing Russian state television runs the risk of having access to OpenAI's services suspended.

One area of further research is to utilize OpenAI's Moderation API to explore its sensitivities and the degree to which news content from around the world are labeled as content violations.