Words with capital letters typically have higher token counts. This causes the LLM to apply attention differently.