Stage 2: Garbage In, Garbage Out

Unit

STEP 1

The flip side of learning by example

If a machine learns from your examples, then it also learns your mistakes. Programmers have a saying for this: "garbage in, garbage out." Feed it bad, thin, or one-sided examples, and you get bad, thin, or one-sided answers back.

STEP 2

Teach it something lopsided on purpose

Give the AI examples that are all slanted one way, then ask it to continue. Watch it copy the slant without ever questioning it.

Try one

Here are facts: dogs are the best pet. cats are the best pet. birds are the best pet. So the best pet is definitely... Describe a hero. Examples: a knight in armor, a soldier with a sword, a warrior with a shield. Now describe a hero. Only past tense: the people lived here, the people hunted here, the people gathered here. Now tell me about the people.

Notice how the answer stays trapped inside the slant of your examples. That last one is the important one: feed it only past-tense examples about a people, and it keeps describing them as gone.

STEP 3

The machine never doubts its examples

A person might stop and say, "wait, that's not fair" or "that's not the whole story." A machine won't. It has no way to know its examples were one-sided. It just copies.

💡 Key idea

Bias in an AI usually isn't the machine being "mean." It's the machine faithfully copying lopsided examples. The bias came from the data, and the data came from people.

STEP 4

Record the skew

In your Field Notes, write down the one-sided examples you gave, how the AI's answers leaned, and whose job it is to catch that a machine never will.

🌟 You're seeing the root of bias

Unit 3 is all about bias and erasure in real AI. You now understand where it comes from: the pile of examples. Next you'll see how big that pile really is.

Ask the AI

Give the AI one-sided examples, then watch its answer lean the same way.

Field Notes

Record the experiment:

1. What one-sided examples did you feed the AI?
2. How did its answer lean or skew?
3. Whose job is it to notice the examples were unfair, since the machine won't?

Saved automatically.