There was a nice paper recently showing that even people encode the knowledge differently depending on which language they use. According to that paper, people

There was a nice paper recently showing that even people encode the knowledge differently depending on which language they use. According to that paper, people made decisions more emotionally in their native languages compared to their second language - where they tend to be more logical. In case of LLMs there were some nice cases where GPT gave different replies depending on the language of the query. Factual information was roughly the same, but in some cases the model gave totally different replies depending on the language of the query. As a bilingual speaker - my experience is that the replies are very similar regardless of whether I speak in Polish or English, or mix them both within the same sentence.