Figurative language, like idioms and similes, adds color and depth to human communication. But can AI truly understand these nuanced expressions? A new research paper explores how different large language models (LLMs), both open-source (like Llama and Qwen) and closed-source (like GPT and Gemini), interpret idioms and similes across multiple languages, including a newly created Persian dataset. The researchers tested various prompting techniques, including zero-shot, one-shot, and chain-of-thought prompting, finding that while these methods can improve AI’s understanding, success varies greatly depending on the complexity of the expression, the language itself, and the specific AI model. Interestingly, open-source models often performed on par with their closed-source counterparts, even exceeding them in some languages, challenging the assumption that bigger, proprietary models always reign supreme. However, low-resource languages like Sundanese and Javanese posed a significant hurdle for many of the models, highlighting the ongoing challenge of ensuring AI inclusivity across the globe's linguistic landscape. The study also revealed that current idiom datasets may not be complex enough, as top-tier LLMs achieved near-perfect scores on existing benchmarks. To address this, the researchers introduced more challenging examples with literal alternatives, revealing even these powerful AIs could be tripped up. This research illuminates the complex path toward imbuing AI with genuine cultural and linguistic understanding. Future research could expand to other forms of figurative language, create more challenging datasets, and focus on bridging the gap for low-resource languages, ultimately enabling AI to navigate the intricacies of human expression with greater accuracy and sensitivity.