Examples, labels, and bias

Imagine teaching someone what a "dog" is, but the only photos you ever show them are golden retrievers.

Then one day you point at a chihuahua.

"That's not a dog," they say. "Too small. Wrong color. No floppy golden ears."

They're not stupid. They learned exactly what you taught them. The problem is what you didn't teach them.

An AI doesn't learn the world. It learns the examples someone chose to show it.

One idea: models copy their examples — gaps and all

Almost every AI you've met learned by being shown thousands or millions of examples, each with a label — a little note saying what it is.

• Photos labeled "cat" or "not a cat."
• Emails labeled "spam" or "safe."
• Sentences labeled "happy" or "angry."

The model's whole job is to find the pattern that connects the examples to the labels. So the examples become its entire universe. Three things follow from that:

If the examples…	Then the model…
Only show one kind of thing	Fails on anything that looks different
Carry a human mistake	Copies the mistake and repeats it at scale
Leave a group out	Acts like that group doesn't exist

That last one has a name: bias. Bias isn't the AI being mean. It's the AI faithfully copying a lopsided set of examples.

Do the thing

For each system below, ask one question: "Whose examples trained this, and who got left out?"

1. A face-unlock feature that works great for the engineers who built it, but not for their grandparents.
2. A voice assistant that understands some accents and shrugs at others.
3. A "predict the next word" model trained mostly on text written by adults, answering a 12-year-old.
4. A photo app that tags "wedding" only when it sees a white dress.

Quick check. Every one of these is a gap in the examples, not a bug in the code. The face-unlock saw mostly younger faces. The voice assistant heard mostly certain accents. The word model read mostly adult writing. The photo app learned one culture's version of a wedding. None of them were told the world is bigger than their examples — so to them, it isn't.

Why this matters

You are about to use tools that were trained on somebody's examples, and you will almost never get to see that pile.

• When an AI is confidently wrong about something specific, ask what it never saw. Wrong answers often trace straight back to a gap in the training examples.
• "Works for me" is not "works." A tool that fits the people who built it can quietly fail everyone else — and the builders are the last to notice.

Next lesson, we stop guessing about those gaps and go find them on purpose. You'll take a working AI and hunt for the exact spot where it breaks.