K-Culture GlossaryㅎWhere everyone starts
Hallucination
When AI states plausible-sounding falsehoods as if they were fact. Generative AI's most notorious weakness.
In plain words
Hallucination is when AI presents a plausible-sounding falsehood as fact. It cites papers that don't exist, references court cases that were never decided, and confidently states wrong dates.
The cause lies in how AI itself works. A large language model isn't a "fact-retrieval machine" — it's a "plausible-next-word machine." Plausibility and truth usually overlap, but where they don't, hallucination happens. It's like a student who doesn't know the answer on a test but fills in something that sounds right anyway — because the model was trained to produce a plausible answer rather than leave a blank.
That's why the number one rule of using AI is: always verify anything important against the original source. AI companies are also working to reduce hallucination by adding search (RAG), giving models more time to "think" (reasoning models), or training them to say "I don't know" when they don't. That's why phrases like "reduced hallucination rate" show up so often in new model announcements.
How it shows up in the news
"A lawyer was disciplined for submitting a fake court case fabricated by AI as-is" — this is a classic hallucination incident, and news like this repeats every year.
Try it yourself
- With the chatbot's search feature turned off, ask about a made-up title: "Tell me the plot of the novel 'Night in the Glass Garden' by author Kim Cheol-su."
- If it plausibly invents a plot for a book that doesn't exist — that's hallucination. (If it says "I can't find such a book," that model is well-trained.)
- Turn search back on and ask the same question again. Without evidence, it should switch to saying it doesn't know — a hands-on way to learn why important information needs source verification.
- One step further: in practice, hallucination isn't eliminated like magic — it's blocked in layers. Parts that must never be wrong get locked down with rules ("always cite the document"), AI judgment is layered on top of that, and if uncertainty remains, a stronger model or a human double-checks. You can use the same trick yourself: for important questions, add "tell me if you're not sure," and cross-check the answer by pasting it into a different chatbot.
Hallucination vs. Bias: What's the Difference
Inventing facts that don't exist, Skewing judgment toward one side
Both get lumped together in articles as "the AI got it wrong," but the fixes are opposite. Invented facts shrink when you supply sources, while skew doesn't shrink no matter how much data you add — because the skew is baked into the data itself.
| Aspect | Hallucination | Bias |
|---|---|---|
| What goes wrong | States nonexistent facts convincingly | Judges unfavorably against a particular group |
| How to check | Fact-checking exposes it immediately | Only visible when comparing many cases together |
| Cause | Can't admit not knowing what it doesn't know | The training data was already skewed |
| How to reduce it | Have it search for sources; require citations | Change data composition and evaluation criteria |
| In the news | Fake case law, fabricated citations incidents | Discrimination controversies in hiring, lending, facial recognition |
Rule of thumbIf a false fact was invented, it's hallucination; if the facts are correct but unfavorable to a particular side, it's bias.
See also
Stories using this term
No story has used this term yet. New ones attach here automatically.