Meet AI: Patterns, Predictions, and Decisions

Goal

By the end of this lesson, you can explain a simple AI system as input → pattern/model → output, tell the difference between a prediction and a guarantee, and name evidence you would check before trusting a result.

What does an AI system actually do?

The word AI can sound much bigger than the thing a system is doing.

A spam filter looks at a message and predicts whether it is spam. A recommendation system looks at information about what you watched or clicked and predicts what you may want next. A camera app may look at pixels and predict what object is in the image.

These systems are different, but we can begin with the same three questions:

1. What information goes in?
2. What result comes out?
3. What pattern or model connects the two?

That gives us a useful first picture:

input → pattern/model → output

Here, a model is the reusable part that turns input information into an output using learned or chosen settings. You do not need to know how those settings are learned yet. For now, think of the model as the system's reusable prediction rule—not the whole app and not a guarantee that the answer is correct.

Patterns are useful regularities

A pattern is something that repeats in a useful way.

Imagine a garden log:

Hours of sunlight	Plant height after two weeks
1	2 cm
2	4 cm
3	6 cm
4	8 cm

In this tiny example, more sunlight is connected with more growth. A simple pattern would be:

about 2 cm of growth for each hour of sunlight

If another plant gets 5 hours of sunlight, a simple system using this pattern might predict about 10 cm of growth.

Learning traceWhat is the predictor actually doing?

Input5 hours of sunlight

→

Transformuse the observed ×2 cm/hour pattern

→

Outputpredict about 10 cm

The 10 cm value is new. The ×2 relationship came from the examples above, so the output depends on both the new input and the pattern we chose to use.

That prediction can be useful even though it is not a promise. Real plants also depend on water, soil, temperature, disease, and many other things that are missing from our tiny table.

This gives us an important habit for the rest of the course:

A pattern can support a prediction without guaranteeing the answer.

Prediction and decision are not the same thing

A prediction estimates an answer that we do not know yet.

A decision chooses an action, sometimes using a prediction.

For example:

System	Possible prediction	Possible decision
Spam filter	“92% likely to be spam”	Put the message in the spam folder
Weather app	“High chance of rain”	Show a rain warning
Recommendation system	“You may like this video”	Put the video near the top of the page

The prediction and the decision are connected, but they are not identical. A product designer still has to decide what action should follow a prediction and how cautious that action should be.

A common mistake: trusting the word “AI”

Suppose two systems both say they use AI.

One was tested on thousands of examples similar to the things it will actually see. The other was tested on only ten easy examples. The label “AI” tells you almost nothing about which system deserves more trust.

Better questions are:

• What kinds of examples was it tested on?
• How often was it correct?
• What kinds of mistakes did it make?
• Is the thing we are asking it to predict similar to the examples it learned from or was tested on?
• What happens if the prediction is wrong?

Trust should come from evidence, not from the name of the technology.

Guided Lab: change one input and read one output

The Lab below uses a tiny Python program. You do not need to understand every line yet. This first activity is only about connecting an input change to an output change.

First, click Run without editing anything.

Look for these two parts of the output:

• new input: 5
• prediction: 10.0

The program found the same “2 output units per input unit” pattern we saw above.

Now make exactly one change:

1. In the Lab below, find the line new_input = 5.
2. Change only the 5 to 6, so the line becomes new_input = 6.
3. Click Run again.
4. Look at the final prediction: value.

Because the examples still follow the same 2-to-1 pattern, the prediction should move from 10.0 to 12.0.

Loading lab…

What happened?

The examples did not change. The pattern estimated from those examples did not change. Only the new input changed.

So the output changed in a predictable direction.

This is a very small system, but the habit matters: change one thing, observe the result, and explain why it changed. Later the patterns will become more complicated, but this habit will stay useful.

What if a new input is very different from the examples?

Imagine that every plant in our garden table was the same kind of houseplant grown indoors. Then we try to use the same model for a cactus growing outdoors in a desert.

The cactus is very different from the examples that produced and tested the pattern. We should be more cautious about the prediction.

This does not prove that the model is wrong. It means we have less evidence that the old pattern works well for this different input.

That idea will return many times in machine learning: examples matter, and predictions on very different inputs deserve extra checking.

Quick Check

1. What does a prediction do?

Chooses the action the system should take after seeing a model outputEstimates an unknown answer from available informationReports a target that is already known in the labeled data

2. A spam system predicts that a message is spam. Which statement describes a decision rather than a prediction?

The message is likely to be spamMove the message to the spam folder based on that estimateThe model used message text as input

3. A plant-growth model was built and tested only with indoor houseplants. Now you want to use it for an outdoor desert cactus. What is the safest first response?

Use the model immediately because it worked on the houseplant dataTreat the cactus as equivalent because both examples are plantsBe more cautious and check whether the old evidence applies

0 of 3 questions answered.

One idea to carry forward

When you meet a new AI system, try to describe it without using the word “AI” at first.

Ask:

What goes in? What comes out? What pattern or model connects them? What evidence would make me trust the result?

That simple description is more useful than treating the system like a black box.

Key Takeaways

• AI systems often use information and learned or designed patterns to produce predictions or decisions.
• A model is the reusable part that maps input information to an output using learned or chosen settings.
• A prediction is an estimate, not a guarantee.
• A decision is an action and may use a prediction as one input.
• Inputs that differ strongly from known examples deserve extra caution.
• Trust should come from evidence about behavior and mistakes, not from the word “AI.”

Next Lesson

Next, you will separate three ideas that are easy to mix up: a hand-written rule, one example, and a dataset made of many examples.

References

• Google for Developers, Introduction to Machine Learning.