Research finds chatbots moderate sensitive topics differently by country

In short: Evidence is mixed on whether chatbots push “moderate” views, and new research instead shows their safety filters vary widely by location and topic.

What's going on

Some commentary has suggested that large language models, often called LLMs (chatbots trained on huge amounts of text), encourage expert consensus and more moderate opinions compared with social media.

But the sources available here do not directly support that broad claim. Instead, they point to a different and more measurable issue, how much chatbots block, refuse, or water down answers on sensitive questions.

A large study presented at the NDSS Symposium 2026 looked at more than 700,000 responses from 15 leading LLMs. The researchers tested 1,118 sensitive prompts across 12 locations and 13 languages, including categories like hate speech, sexual content, politics, and religion. They found big geographic differences. In some settings, “soft moderation” went from about 14.3% to 24.9% depending on language context, meaning the bot would dodge the question or answer indirectly (like a customer service agent giving a vague reply).

The study also found that some topic areas were treated more strictly than others. Hate speech and sexual content were moderated more than political or religious topics, and political and religious answers varied the most by location. Moderated replies were also around half the length of normal replies.

What to watch

If moderation varies by country and language, people may not get the same information or the same level of detail from the same tool. Future research may test whether these differences shape users’ opinions, or whether chatbots actually guide people toward consensus more than social platforms do.

Source: Financial Times