LuWhat is Hallucination?
You ask an AI: “Who was the first person on the Moon?” It says Armstrong—fine. Then you ask: “Who was the first person to eat pizza on Mars?” It confidently gives you a name, a story, and even a date. You know it’s BS, but the tone is dead serious.
That’s “hallucination” in AI—not the drug kind, but when a large model invents stuff with total conviction, without realizing it has no clue.
Why? Because LLMs are fundamentally “word‑chain” machines. They don’t understand “true” or “false”; they only know “statistically, this word is likely to be followed by that.” If your question isn’t clearly covered in their training data, they start a fill‑in‑the‑blank game, stitching together fragments they’ve seen and spitting out something fluent. They have zero ability to distinguish fact from fabrication.
This isn’t a bug—it’s a feature, because they survive on probabilities, not facts. As long as they generate via statistics, hallucinations will never go away entirely.
But hallucinations aren’t all bad. For sci‑fi writing, brainstorming, or comedy, creative BS is gold. But for fact‑checking, legal docs, or medical advice, it’s a ticking bomb.
How to tame it? Common tricks: tighten context (don’t let it roam), RAG (make it retrieve sources first), and good old human fact‑checking (dull but safe). Ultimately, treat AI like a glib intern, not an encyclopedia.
Comments