How Artificial Intelligence Works: Grades 6-9 Without Hype

Quick answer: A machine learning system is given many examples with labels, finds statistical patterns in them, and uses those patterns to guess a label for something new. It gets the guess right often and wrong sometimes, and it has no idea which. Teach it with index cards before you teach it with a chatbot, because the cards make the mechanism visible.

What is the smallest true explanation students can hold onto?

Four words: examples, patterns, prediction, error.

Examples are the training data. Someone collected thousands or millions of items and attached a label to each: this photo is a cat, this email is spam, this review is positive. The labels usually came from people, and who those people were matters later.

Patterns are what the system extracts. It measures features that correlate with each label. Not meaning, not the idea of a cat, just numbers that tend to go together.

Prediction is what happens with new input. The system compares the new item to the patterns and outputs the label that fits best, plus a confidence number.

Error is the part usually skipped. Every system gets some predictions wrong, and the training process works by measuring how wrong and nudging the numbers. Students should learn the word from the start, because a system that is right 95 percent of the time is wrong once in every twenty tries, and in a school of 800 people that is 40 mistakes.

What this explanation deliberately leaves out: understanding. The system is not reasoning about cats. Saying so is not a warning, it is a description of the mechanism. When students later use a chatbot that produces confident, fluent, incorrect sentences, they will already have the explanation for it.

How do you teach it with index cards?

Prepare forty cards. Each has three simple features written on it: length in centimeters, number of legs, and whether it has fur. Twenty are labeled DOG, twenty are labeled CAT. Make the data messy on purpose. Include a small dog and a large cat.

Split the class into groups and give each group thirty cards, holding back ten. Their job is to write a rule that separates the two labels using the features. They may not use anything else. They will start with something like "if it is longer than 50 centimeters it is a dog."

Then test. Read out the ten held-back cards without labels. Groups apply their rule and record a prediction. Reveal the answers and count errors. Every group gets some wrong.

Now the good part. Let them revise the rule using what they learned from the ten. Test again on ten more cards you had also held back. Most groups improve on the second set and get worse on a third, and that is overfitting, taught by experience rather than definition.

How do you handle the training data question honestly?

Rig the cards. In one class set, make every DOG card also say "photo taken outdoors" and every CAT card say "photo taken indoors." Groups will discover this shortcut within four minutes and their rule will become "outdoors means dog."

Then hand them a card that says indoors, 60 centimeters, four legs, fur, labeled DOG. Their rule fails. Ask what went wrong. The rule was not wrong about the data it saw. The data was not representative of the world.

That is the entire bias conversation, grounded in something they built themselves. It avoids both of the bad framings: that these systems are neutral, and that they are malicious. They reflect what they were shown, and someone chose what to show them.

Follow it with a concrete example students can check: two people typing the same query into the same search or video app and getting different results. The mechanism there is a recommendation system trained on what each of them clicked before.

How do you differentiate?

Approaching: two features instead of three, cleanly separable, and a rule template to fill in. "If ____ is greater than ____, predict ____." Their task is the testing and error count, which is where the concept lives.

On level: three features, messy data, two rounds of testing, and a written explanation of why the rule failed on specific cards.

Above level: give them a fourth feature that is pure noise, such as a randomly assigned ID number, and see whether their rule latches onto it. Then ask them to design a test that would catch that mistake. Students who follow this well are ready to look at how recommendation systems shape what people see, which is the subject of Recommendation Algorithms and Filter Bubbles.

What language should you avoid, and what should you say instead?

Avoid saying the system knows, understands, thinks, wants or decides. Say it predicts, classifies, outputs, scores or ranks. Students copy your verbs into their writing, and the verbs carry the concept.

Also avoid the two hype poles. "It will take all the jobs" and "it is just autocomplete" are both conversation stoppers. The accurate middle is that these systems are good at tasks with lots of labeled examples and clear feedback, unreliable where examples are scarce or the right answer is contested, and they produce output with the same confidence either way.

Frequently asked questions

Do I need to explain neural networks?

Not in grades 6 to 9. Rules over features is a truthful simplification. Neural networks learn the features rather than being given them, and that sentence is enough for a curious student.

Should students use a chatbot during the unit?

After the card activity, not before, and check your district policy first. Used first, it teaches wonder. Used second, it teaches analysis.

How long does this take?

Two periods for the card sequence, a third for the biased data set and discussion.

What if a student asks whether it is conscious?

Answer with the mechanism: it turns input numbers into output numbers based on patterns from training. Then say honestly that people disagree about what consciousness would even require, and that the mechanism as described gives no reason to think it is present.

Related posts

• Binary and Data Representation, Unplugged
• Media Literacy: Teaching Students to Check a Source
• Internet Safety Lessons Teenagers Don't Roll Their Eyes At