Module 3: Knowing what not to trust — Knowing what NOT to trust is huge. We'll learn how to spot when AI makes things up.
- Match each hallucination type to its real-world example.
- — Think about what makes each type unique: is the AI inventing, mixing, outdated, or reasoning badly?
- Match each type of truth to the example that best illustrates it.
- — Ask yourself: can this be verified with a clear right/wrong answer, or is it inherently limited?
- Your AI cites a Harvard study to support its answer, but you can't find it anywhere on Google. What type of hallucination is this?
- Fabrication — the AI invented a source that doesn't exist. — When the source itself doesn't exist at all, that's only one category.
- You ask the AI to write a 'strong professional bio' and notice it defaults to traditionally male-coded language. What is the best way to fix this?
- Add explicit constraints to the prompt, like 'use gender-neutral language.' — You can't fix the training data, but you can control the instructions you give.
- Which of the following best describes 'Proxy Data' as a type of truth?
- Information the AI has to trust from human sources because it can't verify it firsthand. — Think about Napoleon's height — the AI can't go back in time to measure him.
- What is the simplest follow-up prompt you can use to catch hallucinations?
- Ask the AI: 'Are you sure? Double-check that.' — Sometimes the simplest check is just asking the AI to review its own work.
- After this lesson, what is the best way to think about AI?
- A smart friend who sometimes makes mistakes and is getting smarter over time. — You wouldn't expect your smartest friend to be right 100% of the time either.
- An AI gives you a confident-sounding answer about what two coworkers discussed in a private meeting. What type of truth problem is this?
- Hidden Data — no recordings or records exist, so the AI literally cannot know. — If no data about the event exists anywhere online, how could the AI possibly know?