Uncertainty is a fundamental part of human communication. We use words like "maybe" or "probably" to express our level of confidence in something, and these words of estimative probability (WEPs) add nuance and depth to our conversations. But what about artificial intelligence? Can AI systems, particularly large language models (LLMs), truly understand and utilize these expressions of uncertainty? A new study delves into this question, examining how LLMs like GPT-3.5, GPT-4, and ERNIE-4 compare to humans in their interpretation and use of WEPs. The researchers found a mixed bag. While LLMs generally align with human estimates at the extremes of certainty (like "almost certain" or "almost no chance"), there's a significant gap in the middle ground. For many WEPs, the LLMs' numerical probability assignments diverged from human judgment. Interestingly, GPT-3.5 often performed more like a human than the more advanced GPT-4, suggesting that bigger isn't always better when it comes to mimicking human-like uncertainty. The study also explored how factors like gender and language affect LLM estimations. Adding gendered pronouns to prompts led to less variability in LLM responses, and while switching between English and Chinese didn't drastically change the GPT models' estimations, there were notable differences between GPT models and ERNIE-4 (a Chinese-trained LLM) when prompted in Chinese. Beyond simply interpreting WEPs, the researchers also tested how well GPT-4 could connect statistical uncertainty (like the probability of a coin flip) to estimative uncertainty. While GPT-4 performed better than random chance, it still struggled to consistently map numerical probabilities to the appropriate WEPs. This suggests that even advanced LLMs have a way to go before they can fully grasp the nuances of uncertainty as humans do. The research highlights the ongoing challenge of aligning AI with human communication, particularly in areas as subtle and complex as expressing uncertainty. As LLMs become increasingly integrated into our lives, their ability to understand and use WEPs effectively will be crucial for seamless human-AI interaction. Future research could explore how these models handle uncertainty in more dynamic conversational settings and with a wider range of probabilistic expressions, paving the way for more human-like AI communication.