Why does AI sometimes make up facts or give wrong answers with total confidence?

AI makes up facts — what researchers call 'hallucination' — because language models don't actually know things the way you and I do. They predict words that sound right based on patterns in their training data, without any real understanding of truth. It's like someone who's read every book in the library but never left the building. They can describe Paris in gorgeous detail — the cobblestone streets, the smell of fresh baguettes, the way light hits the Seine at sunset. But they've never been there. They're remixing descriptions they've read, not recalling an experience. Sometimes they blend two real things into something that never existed. A concrete example: I once asked an AI to summarize a research paper I'd written. It confidently described findings that weren't in my paper at all. The summary sounded completely plausible — it used my actual terminology, referenced real concepts from my field, and matched the tone of my writing. But the specific data points were fabricated. The model wasn't lying. It was just doing what it's built to do: predict the next plausible word. And for a research summary, the most plausible next word is often a statistic or finding, even if the model has to invent one. This is why fact-checking AI output is non-negotiable, especially for anything you publish or use to make decisions. A helpful mental model: treat AI like a very smart, very well-read intern who works incredibly fast but sometimes gets overconfident and doesn't tell you when they're guessing. You still need to verify their work. **Related**: How can I tell if AI-generated text is accurate? | Do all AI chatbots hallucinate at the same rate?