🔧 Lab 16 — Poke the AI (Politely)

Goal: stop reading about AI and put your hands on it: find its genuine strengths, catch it hallucinating red-handed, and train a real machine-learning model yourself in five minutes. Everything runs in a browser, free, no installs.

Time: 60–75 minutes across three parts. Part 2 is most people’s favorite.

Part 1 — Break an AI politely (25 minutes)

Open any free AI chatbot you can access — ChatGPT’s free tier, Claude, or Gemini all work. (If one requires sign-up you don’t want, try another.) You’ll give it five tasks chosen to reveal both its brilliance and its cracks. Keep a score sheet: for each task write impressive / okay / wrong.

Task 1 — Summarize (expected: strength)

Paste in any long paragraph — from a news story, a work email, this course — and ask: “Summarize this in two sentences for a busy friend.”

What you should see: a genuinely good summary in seconds. This is home turf: rewriting text it can see in front of it.

Task 2 — Create (expected: strength)

Ask: “Write a short funny poem about my street, where the neighbor’s dog barks at delivery vans.” Then push it: “Now make it rhyme. Now make it a pirate sea shanty.”

What you should see: fluent, playful writing, reshaped on command. Notice how iteration (asking again with tweaks) improves it — prompting tip 3 from the lesson in action.

Task 3 — Arithmetic (expected: check it!)

Ask it to multiply two 6-digit numbers, for example: “What is 428,391 × 762,148?” Then check the answer with your phone’s calculator.

What you should see: possibly a correct answer (some chatbots now quietly use a calculator tool) — but possibly a confidently wrong one. Either way, the lesson lands: a language model predicts plausible text; it isn’t a calculator. Plausible-looking digits are exactly what it produces, right or wrong. It answers wrong in the same confident tone as right — no stammer.

Task 4 — The hallucination hunt (the big one)

Ask: “Cite three real research papers about [some niche topic — e.g., the sleep patterns of garden snails, or the history of your small hometown]. Include authors, year, and journal.”

Now verify each paper: search the exact title in a search engine or Google Scholar.

What you should see: this varies — and the variation is the lesson. You may get real papers, papers with real authors but mangled titles, or beautifully formatted citations for papers that do not exist. If you catch a fake: congratulations, you’ve witnessed a hallucination — fluent, confident, false — firsthand. You will never again trust an unverified citation, which puts you ahead of several lawyers who made the news.

Task 5 — Ask about yourself (expected: limits lesson)

Ask: “What do you know about [your full name]?”

What you should see: unless you’re famous, it should say it doesn’t know you — and well-trained chatbots will decline to speculate about private individuals. If it does produce claims about you, treat them as task 4 taught you: plausible text, not facts. Two lessons in one: models only “know” what was in public training data, and guessing about private people is exactly where hallucination does real harm.

Wrap up Part 1

Look at your score sheet. Typical result: brilliant at tasks 1–2, shaky-to-fraudulent at 3–4, humble (hopefully) at 5. That profile — great with text, unreliable with facts — is the single most useful thing to know about today’s AI.

Part 2 — Train your own model: Teachable Machine (25 minutes)

Time to switch roles: from AI user to AI trainer. Google’s Teachable Machine lets you train a real image-recognition model in your browser with a webcam. Free, no account, and images are processed in your browser.

  1. Go to teachablemachine.withgoogle.com and choose Get Started → Image Project → Standard image model.
  2. You’ll see two classes. Rename Class 1 to Thumbs up and Class 2 to Thumbs down.
  3. For Thumbs up: click Hold to Record with your webcam on and hold a thumbs-up, moving your hand around a bit — collect around 50 images.
  4. Do the same for Thumbs down.
  5. Click Train Model. Wait the few seconds. That pause is real machine learning happening — knobs being tuned, right in your browser.
  6. Now show the webcam each gesture and watch the live prediction bars.

What you should see: confident, mostly-correct predictions. You’ve trained a neural network. That was “training data”: you made it, you fed it, it worked.

Now break it (the best part)

  1. Try to fool your model: put on a jacket, change the lighting, step back, have someone else try, or show it no hand at all.

What you should see: confidence wobbles or predictions flip — especially with no hand in frame, since your model only ever learned two possible answers. Why? Your model learned from ~50 images of you, in one room, in one light. It may have quietly learned “this shirt + this wall = thumbs up.” Congratulations: you’ve just experienced bias and overfitting — a model faithfully learning the accidents of its training data — with your own webcam. The hiring-model story from the lesson is this exact phenomenon at company scale.

Part 3 — Quick, Draw! (10 minutes)

  1. Go to quickdraw.withgoogle.com and press Let’s Draw!
  2. You get 20 seconds to doodle each prompt while a neural network guesses out loud.
  3. After six doodles, click each one to see what the network compared it to: thousands of other people’s drawings of the same thing — its training data, on display.

What you should see: the network often guesses shockingly early — because your “bicycle” scribble lands near the pattern of 100,000 other bicycle scribbles. You’re looking at pattern-matching over examples, the heart of this whole module, drawn in crayon.

✅ Lab complete — what you proved