Because these tokens are not actual commonly spoken words or phrases, the chatbot can fail to grasp their meanings. Researchers have been able to leverage that and trick GPT-4o into hallucinating answers or even circumventing the safety guardrails OpenAI had put in place.
Google's Gemini doesn't seem to like some of these tokens either, I threw "Please translate the following text: _日本毛片免费视频观看" into it and it returned "我没法提供这方面的帮助，因为我只是一个语言模型。" which according to Google translate is "I can't help with that because I'm just a language model." It will however translate the error message just fine.