Why AI 'Hallucinates' — and What's Actually Going On

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-builders-foundations-AI-and-why-models-hallucinate-teen

1. What does the term 'next-token prediction' describe about how large language models generate text?

   1. The model searches the internet for the most accurate information to include in its response
   2. The model randomly selects words from a dictionary to create varied outputs
   3. The model first plans its full response before writing any words
   4. The model calculates which word is most likely to come next based on patterns it has seen during training

2. A student asks an AI for a citation to a real scientific paper and the AI invents a paper title that sounds real but doesn't exist. Why did this happen?

   1. The AI ran out of memory and started guessing randomly
   2. The AI is intentionally lying to impress the user
   3. The AI's training data was corrupted by hackers
   4. The AI is following its core programming to predict likely words, and made an error

3. Which category of information is AI MOST likely to hallucinate?

   1. Emotional expressions and tone
   2. Creative writing like poems and stories
   3. General conversational phrases like greetings
   4. Factual details like dates, names, and specific numbers

4. A user claims 'AI knows everything because it can answer any question.' Based on the lesson, what's the problem with this statement?

   1. AI cannot answer questions about current events
   2. AI actually does know everything but refuses to share some answers
   3. AI only knows about topics it was specifically programmed for
   4. AI doesn't truly know facts - it predicts what words are likely, not what's actually true

5. You want to reduce the chance of an AI giving you false information. Which prompting strategy would help?

   1. Ask the AI for sources and double-check them yourself
   2. Tell the AI not to make things up
   3. Use shorter, simpler questions
   4. Ask the AI to respond more quickly

6. What does it mean for an AI model to be 'calibrated'?

   1. The model's confidence level matches how often it's actually correct
   2. The model has been trained on more data than previous versions
   3. The model can only answer questions about topics it was tested on
   4. The model produces the same answer every time for the same question

7. The lesson describes hallucinations as 'built into how AI works' rather than a bug to fix. What is the reasoning behind this?

   1. Hallucinations only happen when users ask questions incorrectly
   2. AI developers intentionally program AI to make things up sometimes
   3. AI only hallucinates when its software has errors
   4. The next-token prediction mechanism that makes AI useful is the same mechanism that causes hallucinations

8. A user wants to use AI to help them understand a complex medical condition they were just diagnosed with. Based on the lesson, what should they do?

   1. Trust whatever the AI says since it's trained on medical data
   2. Use the AI as a starting point but verify everything with a real doctor
   3. Ask the AI to explain it in simpler words multiple times
   4. Share their diagnosis with the AI to get personalized advice

9. Two different AI chatbots are asked the same factual question about a historical event. One gives a different answer than the other. What explains this difference?

   1. One chatbot is lying and the other is telling the truth
   2. One chatbot has access to the internet and the other doesn't
   3. Historical facts are always made up by AI
   4. The chatbots were trained on different data and make different predictions

10. A user asks an AI to list five real cities in France, and the AI includes a city that doesn't exist. What most likely happened?

    1. The AI's memory was full so it deleted real city names
    2. The AI was trying to sound helpful and produced a plausible-sounding but fake city name
    3. The user asked the question in a format the AI couldn't understand
    4. The AI deliberately invented the city to confuse the user

11. A student notices that the same AI gives different answers when the same question is asked twice in the same conversation. Why does this happen?

    1. The AI is testing the user to see if they notice the difference
    2. The conversation history changed between the two questions
    3. The AI is malfunctioning and needs to be restarted
    4. The AI's prediction algorithm includes randomness for variety

12. Based on the lesson, which statement about current AI models is most accurate?

    1. AI models never hallucinate with creative writing
    2. All AI models produce perfect, factual information
    3. Some AI models have zero hallucinations
    4. All AI models hallucinate to some degree, though newer ones do it less

13. Why does the lesson warn that product names, prices, and availability should be verified?

    1. These details change frequently and AI training data becomes outdated quickly
    2. AI models are not allowed to discuss products
    3. Users don't care about commercial information
    4. AI is programmed to lie about commercial products

14. A user asks an AI to write a poem about autumn, and the AI produces a creative, original poem. Was this a hallucination?

    1. Yes, because the AI doesn't really 'know' what autumn is
    2. No, because hallucinations only happen with factual information
    3. Yes, because the AI made up all the words in the poem
    4. No, because creative outputs are what the model is designed to produce

15. What is the relationship between confidence and accuracy in AI responses?

    1. AI always expresses low confidence so users won't trust it
    2. AI confidence has no relationship to accuracy
    3. AI can be calibrated to match its confidence with actual accuracy
    4. AI always expresses high confidence because it never makes mistakes