SparkLore
AI writingAdded

Hallucination

할루시네이션

한국어로 읽기

Hallucination is an AI model presenting invented content as fact; in fiction it appears as characters, backstory or rules that are not in the story bible.

Also written as: 환각 · AI 환각 · 설정 왜곡 · confabulation · 지어내기 · 사실 오류

Meaning

Hallucination is a model producing unfounded content with confidence. In factual questions it appears as nonexistent papers or wrong dates. In fiction it takes a different shape. A side character not in the bible appears. The protagonist acquires a younger sister she never had. A magic rule set in chapter 3 quietly changes in chapter 40. A character knows something they could not know. Models are trained to continue a story plausibly, so wherever there is a gap they fill it with something plausible. When that filling contradicts the work's facts, it is hallucination.

Why it matters

In fiction, hallucination appears not as factual error but as continuity error, and readers catch it instantly. One comment, "wasn't she an only child?", erodes trust in the work. The cause is usually the context window: the model did not see the bible (not included), saw it but truncated (over the limit), or saw it but ignored the entry as irrelevant to the scene. Three countermeasures. First, excerpt the characters, places and rules for this scene from the bible into the prompt. Second, state in the prompt: do not invent characters or facts not in the bible. Third, add a review step that checks output against the bible. SparkLore's consistency check does the third. Hallucination cannot be eliminated entirely; the realistic goal is a structure that catches it.

Example

Bible: Han Yuri — only child. Parents died in an accident when she was five. Raised by her grandmother.

Prompt (bible not included): A scene of Yuri talking with relatives at the funeral.

Output excerpt: "Sis, what are you going to do with Grandma's things?" her younger sister Yuna asked. Yuri did not answer.

Yuna does not exist. The request "relatives at a funeral" required family, and with no bible the model made a sister. Adding the bible excerpt ("only child, parents deceased, raised by grandmother") to the same prompt turned the scene into an awkward exchange with a distant relative, which fit the character.

Common mistakes

  • Generating without the bible. The model fills gaps. Do not leave gaps.
  • Pasting output without reading it. Skim for characters, places and facts before it goes into the manuscript. Thirty seconds.
  • Blaming only the model. Most of the time the information was not in the prompt. Fix the context.
More AI writing terms