🌀 Machines That Talk · Day 9

Why Bots Make Stuff Up

A chatbot's real job is to write text that sounds right — not text that is right. So when it doesn't know something, it can just... confidently invent it. Let's catch a bot in the act.

↓ ask the bot
📚 Lab · The Knowledge Shelf

Ask the bot. Watch its confidence.

The bot only truly learned these 6 facts. Pick a question below. Some are on the shelf — some are about made-up planets and future events. Then flip 🎲 Guess mode and see what changes.

📚 that's everything it actually knows. Now pick a question:

honest — admits when it doesn't know
🤖 Pixel saysPick a question to hear my answer.
Confidence

Notice the confidence meter — the scary part is when it stays high even when the bot is dead wrong.

🐞 Break-It Challenge

Make it invent with a straight face

Tap a button — it flips Guess mode ON and asks the bot about something that does not exist. Read its answer up in the lab. Full sentences. Specific numbers. Total nonsense.

It's not lying on purpose. The bot has no "I'm unsure" alarm — it just pattern-completes: it fills in words that fit the shape of an answer, whether or not they're true. That's why you always verify important facts before trusting them.
🏷️ The pro wordA confident answer that's just made up is called a hallucination. When a model instead shows honest doubt — "I'm not sure" — that's calibrated uncertainty. A good model's confidence should match how likely it is to actually be right.
🕵️ Boss Build · Spot the Hallucination

Train your skeptic sense

The bot gives an answer — sometimes true, sometimes invented. You be the fact-checker. Vote 🟢 Real or 🔴 Made-up, then reveal the truth. Catch 3 to win the badge!

Caught: 0 / 3
🚀 Beyond the Basics

How the pros fight hallucination

A raw model just guesses. Here's the ladder from "makes stuff up" to "actually trustworthy":

🎲

Pure guessing (a raw model)

Always produces a fluent answer, even about things it never learned. ✅ Never stuck for words. ⚠️ Confidently wrong — the classic hallucination.

🤔

Teach it to doubt (what you watched)

Train the model to say "I don't know" and show real uncertainty. ✅ Much safer. ⚠️ Can get over-cautious and refuse easy questions.

📎

Ground it in sources (the real deal)

Before answering, fetch real documents and answer from them — with citations you can click. ✅ Checkable. ⚠️ Only as good as the sources it fetched.

🏷️ Pro names to look upMaking stuff up is hallucination; matching confidence to correctness is calibration. Fetching real documents to answer from is retrieval-augmented generation (RAG), and backing each claim with a source is grounding & citations — the big idea we build in Day 16.

🎉 Day 9 Complete!

You learned the bot's secret: it's built to sound right, not to be right. When it doesn't know, it can hallucinate — invent an answer with full confidence. Now you know to check the confident-sounding stuff, every time.

Day 10: The Attention Room →
Tomorrow: how does a bot decide which words to pay attention to when it writes the next one? We step inside the Attention Room. 👀
AI Adventure · Machines That Talk · Day 9 — made for young inventors 🚀