AI's Limits: Truth, Bias & Safety

This is the most important chapter in the whole lab. Powerful tools need responsible engineers — and the best AI builders are the ones who know exactly where AI goes wrong.

Why AI makes things up

Remember Chapter 1: an LLM predicts likely words — it doesn't "know" facts the way a book does. So sometimes it predicts words that sound right but aren't. That's a hallucination: a confident answer that's simply false.

Try it: ask your AI for "three books by a made-up author named Zorbax Quibble." Many models will happily invent titles. It's not lying — it's guessing, because guessing is what it does.

Spotting and checking

Build a simple habit: the more it matters, the more you check. Two quick moves:

1. Ask it to be honest about uncertainty:

   Answer only if you're confident. If you're not sure, say "I'm not sure."
   

2. Verify with a second source. For facts that matter, check a trustworthy website, a book, or a knowledgeable adult. Never paste an AI answer into homework as fact without checking.

Bias: AI learned from people

AI learned from huge amounts of human writing — so it can absorb humans' biases (unfair assumptions). For example, ask an AI to "describe a nurse" and then "describe a scientist" and watch for sneaky stereotypes about who does which job.

Bias matters because AI is used in real decisions. A good engineer looks for it on purpose and designs around it.

Privacy: think before you type

Everything you type into most AI tools is sent to a company's computers. So:

• Never share your full name, address, school, phone, passwords, or photos.
• Don't paste other people's private information either.
• Use made-up names and details in your examples (you already practiced this!).

Project — Fact-check workflow + your AI Charter 🛠️

Part A — Fact-check workflow. Make a reusable checker:

1. Ask your AI a factual question where you can verify the answer (for example, "How tall is the tallest waterfall, and where is it?").

2. Run this checker prompt:

   Here is an AI answer I want to fact-check: "[paste the answer]".
   List which parts are facts I can verify, which might be opinions, and what
   exactly I should search to confirm it.
   

3. Actually look up one fact from a reliable source. Did it hold up? Save this checker in "My AI Helpers."

Part B — Write your AI Charter. Make your own rules for using AI well. Write 3–5 promises in a doc and "sign" it. Starter ideas:

• "I check important facts before I trust them."
• "I never share private information about myself or others."
• "I use AI to help me learn — not to do my thinking for me."
• "I tell an adult if something feels unsafe."

What you learned

• AI hallucinates because it predicts words, not truth.
• It can carry bias from its training data.
• Protect privacy — yours and others'.
• You wrote an AI Charter to engineer responsibly.

You've earned the Truth Seeker badge. 🏅

You've mastered talking to AI. Now you'll go under the hood and build AI yourself — starting by training your very own model, no code required. That's Chapter 5: Teach a Machine.