Unit 4What AI actually is

The big idea

There are two kinds of work a computer can do. Knowing the difference tells you when to trust AI and when to check it.

• Fixed rules (deterministic): the same input always gives the same output. 1 + 1 is always 2. Old software was a giant tree of "if this, then that," with every branch written by hand.
• Best guesses (probabilistic): the answer depends on context. A weather forecast says "90% chance of rain." Picking an outfit depends on the day. This is closer to human intuition, and it's how AI models work.

Good AI systems combine the two. Fixed rules say what the AI can and can't do, and the model reasons inside those rules. Without the rules, AI starts confusing people and making things up.

Why it matters

If you know where AI is likely to be wrong, you know where to spend your attention. This unit is why the rest of the course spends so much time on checking.

The fractions lesson

Say an AI answer is about 80% right. Now you build a second step on top of that answer. Your odds are now:

0.8 × 0.8 = 0.64

You lost 16 points without noticing. Add a third step and you're at about 51%. The longer the chain of unchecked steps, the worse it gets.

What to do about it:

1. Start a new chat for each new problem. Don't let one long conversation pile up mistakes.
2. Put everything in one complete prompt instead of a long back-and-forth. (That's Unit 5.)
3. Check answers in a second AI from a different company. Each model has its own bias, so they catch different mistakes.

Do this

1. Write your own definition of AI in one sentence. In a group, share a few and sort them into "fixed rules" and "best guesses."
2. Do the math above on paper. Then work out what happens after four steps.
3. Ask two different AIs the same factual question about your project. Compare the answers. Where do they disagree?
4. Pick one claim from either answer and check it yourself somewhere else.

AI as a thinking partner

Think of AI like an older sibling. You go to them because they know something you don't. But you still decide what to do.

AI tools aren't really there to give you the answer. They're there to help you think and help you find the answer. You still have to use your judgment.

One more term: a context window is how much information a model can hold at once. A few years ago you had to paste things in small pieces. Now some models can hold a whole company's worth of documents. More context means better answers, but only if what you give it is good.

Questions to ask yourself

• Which parts of my project need a fixed, exact answer? Which parts need judgment?
• How many steps am I chaining together without checking?
• Did I check this answer anywhere other than the AI that gave it to me?

In my words

"If you put garbage into a system, you're going to get garbage out of a system. Always look to question what you're getting."

"Ninety percent of my day is spent deciding where we put humans and where we put AI."

Worksheet

Use Worksheet 4: The fractions check in the PDF kit.