Large language models (LLMs) have shown remarkable abilities, but can they truly understand the facts they learn? New research suggests a surprising limitation called the 'reversal curse.' While LLMs can easily learn that 'A is B,' they often struggle to infer the reverse, that 'B is A.' This seemingly simple problem reveals a deeper issue: LLMs might be relying on the *structure* of information rather than genuine understanding. For example, in a study using biographical data, models trained on sentences like 'Daphne Barrington is the director of A Journey Through Time' could answer related questions. However, when trained on the reversed structure, 'The director of A Journey Through Time is Daphne Barrington,' they failed. This suggests a 'thinking bias' where LLMs prioritize names as the starting point for recalling information. Further investigation using chain-of-thought prompting and attention analysis confirmed this bias. Interestingly, even when given multiple-choice questions where both 'A' and 'B' are present, this bias persists. Efforts to mitigate this bias through longer training, mixed training data, or question-answer fine-tuning have proven largely unsuccessful, highlighting the challenge of teaching LLMs true relational understanding. This research raises important questions about how LLMs learn and generalize, and suggests that simply increasing model size or training data may not be enough to achieve true AI reasoning.