A new study shows large language models find it much harder to understand the nuances of Indian, British and Australian English.