Large language models generate confident-sounding text, which makes it easy to form wrong mental models about what they actually do. Understanding what LLMs are not — not databases, not reasoning engines, not search — is as important as understanding what they are for using them effectively.

Several persistent myths shape how people use LLMs, and each one leads to real failures. Myth one: 'LLMs know things.' No — LLMs predict the most likely next token based on patterns in training data. They don't 'look up' facts; they statistically approximate them. This is why they hallucinate with confidence: generating fluent text and generating correct text are different processes. Myth two: 'Bigger models are always better.' Not true. For most tasks, a well-prompted smaller model outperforms a poorly-prompted large one. Model size matters for hard reasoning, but for summarization, classification, and routine generation, size hits diminishing returns quickly. Myth three: 'LLMs understand language the way humans do.' They don't. They process tokens — sub-word fragments — and their 'understanding' is statistical pattern matching over vast text, not grounded comprehension. This is why LLMs struggle with simple physical reasoning, spatial relationships, and counting. Myth four: 'LLMs reason.' Modern reasoning models do externalize reasoning traces, but even these are pattern-matched plausible reasoning, not logical deduction. Myth five: 'LLMs are objective.' They encode the biases of their training data and fine-tuning feedback. The practical takeaway: treat LLMs as extremely capable text predictors with predictable failure modes, not as omniscient reasoning machines. Use them for what they're good at, verify what matters, and never confuse fluency for accuracy.

What Are Common Misconceptions About LLMs?