Large language models (LLMs) like ChatGPT are increasingly used as sources of information, even on complex topics like climate change. But what happens when these AI models are fed misinformation? New research explores this critical question, examining how LLMs react to false climate claims and testing methods to "unlearn" this inaccurate information. Researchers "poisoned" an LLM by feeding it a dataset of false climate information. Surprisingly, the poisoned model still performed well on other topics, suggesting that AI can hold conflicting knowledge depending on the subject. This raises concerns about targeted misinformation campaigns and the need for robust testing procedures. The study also investigated ways to correct the LLM's inaccurate climate knowledge. They found that "unlearning" the false information was more effective than simply training the model on correct data. This suggests that addressing misinformation directly is crucial for building trustworthy AI. Furthermore, giving the LLM access to accurate information during use, through a technique called Retrieval-Augmented Generation (RAG), also proved effective. This highlights the importance of providing AI with reliable sources. This research underscores the challenges of ensuring AI accuracy in the face of widespread misinformation. As LLMs become more integrated into our lives, safeguarding them against manipulation and ensuring they provide reliable information is paramount. Future research will explore how these findings apply to other AI applications, especially in high-stakes areas like healthcare and energy.