Dog whistles – coded language with hidden meanings – are increasingly used to spread harmful ideologies online, bypassing traditional content moderation. Researchers have developed a novel AI task called FETCH! (Finding Emergent Dog Whistles Through Common Habitats) to identify these emerging coded phrases in large datasets of social media posts. The research explores whether AI can detect these nuanced expressions, which often go unnoticed by human moderators and existing algorithms. FETCH! uses datasets from various platforms like Reddit, Gab, and X (formerly Twitter) to represent different online environments. The initial results show current AI methods, including popular techniques like Word2Vec, BERT, and even large language models, struggle with the complexity of the task. Researchers introduced a new method, EarShot, which combines the power of vector databases and LLMs. EarShot shows promising improvements in detecting these hidden meanings. It first converts social media posts into numerical vectors, creating a map of semantic relationships. It then identifies posts similar to those containing known dog whistles. Finally, EarShot uses keyword extraction or direct prompting of LLMs to uncover potentially coded phrases within these similar posts. While EarShot outperforms other methods, the overall accuracy remains a challenge. A key finding is that AI systems, like their human counterparts, tend to prioritize caution, often overlooking dog whistles to avoid false alarms. This conservative approach, while helpful for reducing the burden on human reviewers, limits the AI's ability to uncover the full extent of coded language. The future of this research lies in developing more sophisticated AI models that can effectively balance precision and recall, catching harmful coded language without raising too many false flags. The goal is to refine systems like EarShot, exploring hybrid approaches that combine keyword extraction with the powerful, yet sometimes overzealous, generative capabilities of LLMs. This work highlights both the potential and the current limitations of AI in tackling the increasingly complex landscape of online hate speech and disguised harmful communication.