Evaluating task understanding through multilingual consistency: A ChatGPT case study