AI

Why do AI models hallucinate?

How AI hallucinations happen: why LLMs predict likely words instead of checking facts, and how to cut made-up answers with verification, grounding and prompts.

  • 7 min read
  • 5 questions
  • +35 XP
Transcript

AI hallucinations, explained by cats. Kitty asks the AI, “Who invented the tuna compiler in 1987? Cite sources.” The AI replies instantly. Very polished. Very confident. Tiny problem: the person isn’t real. The date is made up. The citation is imaginary.

Chaos. That’s an AI hallucination. It sounds true. But the LLM just guessed. Why? Because an LLM does not check truth. It predicts the next likely word. Then the next. Then the next. Usually that works great. But if the prompt is vague, the facts are missing, or the answer was never in its data, it may fill in the

gaps. And because it’s trained to be useful, it may answer the question instead of saying, “I don’t know.” So how does Kitty avoid this? First: verify everything. Names. Dates. Stats. Sources. Second: give the model reference material. If Kitty provides the actual documentation, the AI has less room to invent tuna history.

Third: prompt better. Be specific. Set boundaries. Tell it to say when it is uncertain. AI hallucinations are not lies. They’re prediction mistakes by a model that's made to predict, not tell the truth.