You’ve met three AIs: one that sees, one that reads feelings, one that writes. Today’s project is to test them like a scientist — try them on your own inputs, record what’s impressive and what’s silly, and start asking the most important question about AI: why does it get things wrong?

💡 In Colab. Load all three (each downloads the first time):

!pip install -q transformers
from transformers import pipeline
recognizer = pipeline("image-classification")
mood = pipeline("sentiment-analysis")
writer = pipeline("text-generation", model="distilgpt2")

Run your tests

Try each AI on a few inputs of your choosing:

# feelings
print(mood("I can't wait for summer vacation!"))
print(mood("Ugh, my tablet battery died again."))

# writing
print(writer("My favorite invention would be", max_length=40)[0]["generated_text"])

# seeing (upload a photo first)
from google.colab import files
uploaded = files.upload()
filename = list(uploaded.keys())[0]
print(recognizer(filename)[0])

Keep a lab notebook

Real scientists record results. As you test, jot down (in print statements or on paper) two columns: Clever (it nailed it) and Clueless (it blew it). Hunt for failures on purpose — a sarcastic sentence, a blurry photo, a weird story prompt.

Try it 🎯

1. Find one example where each AI does great.
2. Find one example where each AI is wrong or weird.
3. For one wrong answer, write down why you think it got confused.

Why does AI get things wrong?

Here’s the heart of it. An AI only knows what it learned from its training examples. So:

• It fails on things unlike its examples (a cartoon when it learned from photos; sarcasm when it learned plain reviews).
• It can be confidently wrong — the score is “how sure,” not “how right.”
• It can be biased: if the examples it learned from were lopsided or unfair, its guesses will be too. An AI trained mostly on one kind of photo, or one kind of writing, does worse on everything else.

None of this means AI is bad — it means AI is a mirror of its examples. That’s a powerful idea to carry forward, because next you’ll be the one choosing the examples.

Think about it 🔮

An AI that learned to recognize animals from photos of cats and dogs is shown a photo of a horse. What’s likely to happen? (It’ll probably guess “dog” or some animal it does know, confidently — it can only choose from patterns it learned. It can’t say “I’ve never seen this.”)

Your mission 🚀

Write a short “AI report card.” Test each of the three AIs on three inputs (nine tests total). For each, print the input, the AI’s answer, and a ✓ or ✗ for whether it was right. End with one sentence: what kind of input fooled the AIs most?

What you learned today

• You can combine AIs and test them like a scientist — and finding failures is the goal, not a problem.
• AI only knows its training examples, so it fails on anything unlike them.
• “Confident” ≠ “correct”; AIs can be sure and wrong.
• AIs can be biased if their examples were — they mirror what they learned.

You’ve used AI and seen its limits. Next, we flip the script: instead of using a trained AI, you’ll learn how one learns — starting with a tiny example. 🍎🍊