Cryptic crosswords, those fiendishly clever puzzles loved by wordplay enthusiasts, have long been a challenge for artificial intelligence. While AI models like ChatGPT can translate languages and write poems, they often stumble when faced with the subtle nuances and intricate wordplay of cryptic clues. A new study delves into the reasons behind this struggle, exploring why even large language models (LLMs) can't quite crack the code. Researchers tested popular LLMs like Gemma2, LLaMA3, and ChatGPT on a massive dataset of cryptic clues from The Guardian, and the results reveal a significant gap between human and machine performance. Even with clever prompting techniques, including feeding the AI the definition part of the clue, the models lagged significantly behind human solvers. The study suggests that LLMs struggle with several key aspects of cryptic crossword solving. They have difficulty extracting the definition part of the clue, correctly identifying the type of wordplay used (like anagrams, hidden words, or double meanings), and explaining their reasoning process. Interestingly, the models seem to perform best with double definition clues where both parts of the clue are synonyms of the answer, likely because this aligns more closely with how LLMs are traditionally trained. This research highlights the complexity of true language understanding and reasoning. While LLMs are adept at pattern recognition and statistical analysis, cryptic crosswords demand a deeper level of semantic understanding and the ability to manipulate language creatively. The research team suggest promising future directions include chain-of-thought prompting, where the AI is guided through the reasoning process step-by-step, and curriculum learning, gradually increasing the complexity of the clues the AI is trained on. While AI may not be a cryptic crossword champion yet, this research offers valuable insights into the ongoing quest to build truly intelligent machines capable of complex linguistic reasoning.