Why AI Hallucinates and What Actually Reduces It

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-hallucination-final1-creators

1. Is hallucination a bug to be patched or intrinsic to language models?

   1. A simple bug fixable by a patch
   2. Intrinsic to how language models work
   3. A network glitch
   4. A user error only

2. Which technique meaningfully reduces hallucination?

   1. Asking 'be honest' once
   2. Increasing temperature
   3. Retrieval-augmented generation with explicit citation requirements
   4. Removing all instructions

3. What's a useful instruction for uncertain questions?

   1. Demand confident answers always
   2. Ban the word 'maybe'
   3. Forbid citations
   4. Tell the model to express uncertainty when it lacks confidence

4. Why cross-check facts programmatically?

   1. A program can verify against trusted sources at scale
   2. Programs can read minds
   3. Programs are always wrong
   4. Programs are decorative

5. What does 'fail gracefully' mean in this context?

   1. Crash the app
   2. Design flows that recover when the model is wrong
   3. Hide errors from users
   4. Always trust the first answer

6. Can hallucination be eliminated entirely?

   1. Yes, with the right prompt
   2. Yes, with a paid plan
   3. No — only reduced significantly
   4. Yes, with multilingual prompts

7. Can a model reliably distinguish its own confident-correct from confident-wrong outputs?

   1. Yes, models always know
   2. Only paid models can
   3. Only frontier models can
   4. No — that's why external verification matters

8. Should you trust the model's own reports of its uncertainty?

   1. Treat them as a hint, not ground truth
   2. As gospel
   3. Ignore them entirely
   4. Only trust them in JSON

9. For high-stakes outputs (legal, medical, financial), what's the rule?

   1. Trust freely
   2. Assume it might be wrong and build verification into the workflow
   3. Avoid AI entirely
   4. Print and hope

10. What concrete experiment compares mitigation techniques?

    1. Ask once and trust it
    2. Avoid factual questions
    3. Ask the same factual question three ways: bare, with sources, with 'say I don't know if uncertain'
    4. Use only fiction prompts

11. What is 'grounding' in this context?

    1. Plugging the model into the wall
    2. Disabling the model
    3. Avoiding all output
    4. Tying output to retrievable, citable sources

12. Which is true about hallucination across model sizes?

    1. Even frontier models hallucinate
    2. Only small models hallucinate
    3. Frontier models never hallucinate
    4. All models are equally bad

13. What is calibration?

    1. The screen brightness
    2. How closely a model's expressed confidence tracks reality
    3. The font size
    4. The model temperature

14. What's the danger of 'prompted-away' hallucination?

    1. It actually solves the problem
    2. It increases cost
    3. It feels solved while still happening silently
    4. It disables the model

15. What's the right framing for a designer building on top of an LLM?

    1. Hallucination is a temporary glitch
    2. Hallucination is the user's fault
    3. Hallucination is illegal
    4. Hallucination is a property of the substrate; design around it