What AI Is and How It Learns from Examples

You may use artificial intelligence, or AI, without seeing it. AI can work inside a phone, website, car, or game. Look at the job a tool does to start spotting it.

Listen, read, and suggest

A voice assistant uses AI to understand a spoken question. You might ask, "What's the weather today?" Speech-to-text turns speech into writing. You speak, and words appear on the screen.

Predictive text, also called autocomplete, suggests words while you type. A translation tool changes words into another language. For example, "Where is the library?" becomes "¿Dónde está la biblioteca?" in Spanish.

A recommendation system suggests something you might like. After several shark shows, you might see another sea-animal show suggested. These tools use patterns in examples to make their suggestions.

Recognize and create

A photo app may find your beach pictures using AI. A phone may recognize its owner's face. A drawing game may guess your sketch before you finish.

Other AI tools make something new. A chatbot writes replies to your questions. An image generator can make a picture from "a cat wearing a space helmet." This is generative AI: AI that creates text, pictures, or sounds.

Does a robot always use AI?

A robot body tells you very little about how it works. A toy that follows "forward 2 steps, turn right" follows programmed moves. A vacuum that recognizes socks with its camera uses AI. A translation app has no robot body at all.

An ordinary calculator follows exact rules for each step. It gives 56 for 7 times 8; it did not learn this from example pictures or messages.

Recognition, suggestions, and new creations are useful clues when looking for AI. But a rule-following program can do some of these jobs too. A spell checker may simply compare words with a list. To find out how a tool works, check its makers' description with an adult.

Artificial intelligence is a computer system that does jobs usually needing human thinking. Recognizing a face and understanding speech are examples. AI imitates parts of human intelligence. The word artificial means made by people.

People decide what an AI tool should do. They build and test it. They also choose the examples it learns from. AI is technology people create, even when its answers seem surprising.

Learning from examples

Most everyday AI tools are made using machine learning. People give a computer many examples. A learning program finds patterns, or details that repeat. People do not write a separate sorting rule for every case.

Imagine many pictures labeled "banana." The computer may find long, curved shapes in them. It uses those patterns when shown a new picture. It can still make a mistake, especially when that picture looks unlike its examples.

Two kinds of jobs

A guess about a group and a newly written story are different outputs. Both can come from patterns learned during training.

Friendly words are still words

You understand things through your own life. You have feelings, wishes, and experiences. AI does not think or feel the way you do. It works with learned patterns, without understanding a dog as you do.

A voice assistant might say, "I'm happy to help!" The words sound friendly. They do not describe a feeling inside the computer.

Fast answers can be wrong. A chatbot may sound sure while giving a made-up fact. The speed and tone of an answer cannot tell you whether it is true. Facts that matter need checking against trusted sources, such as a library book or an official website.

An AI system needs several parts working together. Data supplies information. A learning program finds patterns. The model holds the patterns learned. Hardware is the computer equipment that runs the system.

First, make the model

Data can be pictures, sounds, words, or numbers. Examples used for learning are called training data. A speech tool might learn from recordings paired with written words.

The learning program follows an algorithm, a set of steps. It works through the examples during training. Training produces a model. The program does the learning work; the model is the result you use afterward.

Think about making a speech-to-text model. Many recordings show how spoken sounds match written words. The program finds patterns across those pairs. The finished model can then guess words in a new recording.

Next, use the model

When you use a trained model, follow three parts:

1. Input: information goes into the model.
2. Model: learned patterns are used.
3. Output: the model gives a result.

Suppose you say, "Remind me to feed the fish."

• Your recorded voice is the input.
• The model matches sound patterns to words.
• The written reminder on the screen is the output.

A recommendation tool follows the same path. Your earlier video choices go in. The model uses patterns in what viewers chose. A list of suggested videos comes out. That list is a guess about what you may enjoy.

Where the equipment fits

A microphone collects sound. A camera collects pictures. Computer chips run the model. A screen shows text or pictures, and a speaker plays sound.

The camera is not the model. Taking a picture collects information; recognizing something in that picture is a different job.

Some models run on the device you hold. Others run on large computers far away. Your device sends information over the internet and receives the result. In either case, the input goes through a model before you receive its output.

Imagine helping a computer sort fruit pictures. Each picture is an example. The whole collection is a dataset. A label names the group an example belongs to, such as "apple" or "banana." Examples paired with labels are labeled data.

Look at the details

A feature, also called an attribute, is a detail you can notice. Color, shape, size, texture, and sound can be features. Which features help depends on what you want to sort.

Look through these six fruit pictures in order:

First, count the labels. There are three apples and three bananas.

Next, try color. Yellow appears in both groups. Green appears in both groups too. Color alone cannot separate these pictures.

Now try shape. Every apple here is round. Every banana here is long. Shape separates these six examples. The feature "is a fruit" would not help: all six share it.

Give each example one clear home

Labels need to be correct. A banana labeled "apple" teaches the wrong pattern. The computer learns from the label it receives; it cannot spot that person's labeling mistake for you.

Labels also need to stay separate for this sorting job. "Apple" and "red fruit" overlap because a red apple fits both. "Apple" and "banana" give each of these pictures one clear home.

How this differs from your own sorting

You can sort buttons using questions like "Is it round?" A branching database follows the questions someone wrote. With machine learning, you provide labeled examples. The learning program works out patterns for sorting them. You choose the groups and examples, without writing every question the model will use.

You can act out training with picture cards. You can also use a no-code training tool chosen by your adult. A model that sorts examples into groups is a classifier.

Choose the job and labels

Suppose you want to sort hand pictures. Choose "thumbs up" and "thumbs down" as your two labels. The labels tell the model what answers it can give.

Collect and label examples

Collect 30 thumbs-up pictures and 30 thumbs-down pictures. That makes 60 pictures altogether. Put each picture under the correct label.

Make the examples varied. Include different hands, left and right hands, near and far views, and bright and dim light. Seeing many views helps the model learn hand shapes beyond one exact picture.

Train

The tool's learning program looks through the examples. It finds patterns for each label and builds a model. You do not give it the rule "the thumb points up." The training step finds patterns from the pictures.

Try something new

Show a hand picture the model has never learned from. Look at its chosen label. Try several new pictures, including a different person's hand.

Some tools show how sure the model is about a choice. Even a strong-looking result can be wrong. Compare the label with the picture yourself.

Learn from a mistake

A sideways thumb might receive "thumbs up." Look at the labels: neither says "sideways." The model must choose from the labels it has.

You could add a "sideways" label and collect examples for it. Train again, then try new sideways pictures. This gives the model another possible answer. Keep trying and improving as you find problems.

Show the original two-label model a cat, and it still chooses a thumb label. Its answer list has no "cat" or "neither." A label appearing on screen does not mean the picture belongs to that group.

You can follow these steps for other data too. A sound classifier might sort claps and whistles. A pose classifier might sort standing and sitting. Other models make predictions, such as guessing the next word.

After training, a model uses patterns to make a prediction. This is its best guess about something new. It might guess a label, a spoken word, or a book you will enjoy.

Why one word comes first

Predictive text learns which words often appear together. After "Happy," it may suggest "birthday." That pair appears often in writing.

Imagine ten messages beginning "See you." Six continue with "soon," three with "tomorrow," and one with "later." In this simple example, "soon" is the first suggestion because it came next most often.

The model does not know your plans. If you will meet your friend tomorrow, "soon" may be less useful. You can change or ignore the suggestion.

Tone matters too. "Dear Sir" sounds odd in a friendly message to your best friend. A suggestion can follow a pattern while still being wrong for your message.

How recommendations work

A recommendation system looks for patterns in choices. Many people who liked book A may also have liked book B. If you liked book A, it may suggest book B.

That suggestion can be helpful. It can also keep bringing you similar things. You may miss a different kind of book you would love. You can search for something new yourself.

Guessing drawings and spoken words

A drawing game learned from many labeled drawings. It compares the lines you draw with patterns for labels like "cat." Enough matching lines may let it guess before you finish.

Speech-to-text predicts which words match the sounds it receives. Some sounds are hard to tell apart. It may write "ice cream" when you meant "I scream." Fast speech and unusual names can confuse it too. Read the words on the screen and correct them when needed.

What a model recognizes depends on its training examples. In a simple training tool, adding examples changes the model only after you train again. Then you need new examples to find out what improved.

Fill a gap

A fruit model learned from 20 red-apple pictures and 20 yellow-banana pictures. It may get a green apple wrong because it never learned from green apples.

Add ten green-apple pictures under "apple." There are now 30 apple examples. Train again using the changed collection. Then show different green-apple pictures that were not used for training. Their answers help you check whether the model improved.

More is helpful; variety matters too

Two pictures per label give a model very little to learn from. More examples usually help. But repeating almost the same picture leaves many things unseen.

Fifty photos of one apple on one plate show little variety. Twenty photos of different apples, sides, and lighting show more. Choose examples like the things the model will meet later.

Check the labels

Suppose five banana pictures were wrongly labeled "apple." Adding many pictures will not correct those five labels. Fix their labels and train again. Otherwise, the model keeps learning from the mixed-up information.

Notice accidental patterns

Imagine all dog pictures show grass. All cat pictures show rooms. The model might learn "green background means dog." A cat on a lawn could then receive the label "dog."

To untangle the pattern, add cats outdoors and dogs indoors. Train again. Try new pictures of both animals in both places. Check whether the model still seems confused by the background.

If one kind of example keeps going wrong, look for it in the training collection. Is it missing? Is there only one view? Is its label correct? The mistakes help you choose what to change next.

Some tools, including recommendation systems, keep learning from what people click. Their suggestions can change over time. The simple model you train stays the same until you train it again.

A model needs examples of the people and things it will meet. Missing groups can lead to more mistakes for those groups. When a tool works worse for some people, or treats them worse, it is unfair. Unfair results linked to its data or design are called bias.

Count the groups

Training data is balanced when labels have about the same number of examples. It is unbalanced when some labels have many more examples, or a group is missing.

A pet model learns from 40 cat pictures and four rabbit pictures. There are 44 pictures altogether. It sees ten times as many cats as rabbits. It will probably make more mistakes with rabbits because it has so few rabbit examples.

Adding rabbit examples addresses that gap. You must also check their variety. Many pictures of the same rabbit would still leave other rabbits unfamiliar.

Look inside each group

Even equal totals can hide missing examples. A voice tool may have many recordings, but mostly adult voices. Children's voices sound different, so it may understand children less well.

Adding many children's recordings gives the tool more examples of those voices. Include different speakers and ways of speaking too. Try it with new speakers afterward.

An image generator can repeat narrow patterns as well. If its scientist pictures mostly show men in white coats, it may keep drawing that kind of scientist. Real scientists include people of different genders and backgrounds, in different clothes and workplaces.

Training examples can also record unfair choices people made in the past. A model may learn those patterns and repeat the unfairness.

Make a fairer plan

People are responsible for the examples and the tool they build. An AI tool has no plan to be mean. It can still cause unfair results by repeating patterns in its data.

Ask which people or things are missing. Add correct, varied examples that include them. Keep the numbers across labels reasonably balanced. Train again, then try new examples from different groups. Check whether one group still gets more mistakes.

If a tool misses a child’s words, the child should not be blamed for how they speak. The people building the tool need to improve its data and design.

You and a machine learning system both learn from examples. Practice and feedback can help both of you improve. Both people and machines make mistakes. These similarities can make AI seem like a person.

What learning means for you

You make sense of experiences. You ask questions, connect ideas, and try things out. You can explain why you think something happened.

After meeting a few dogs, you may recognize a very different dog. You connect what you see with barking, fur, movement, and memories of playing. You can also recognize a cartoon dog without having met one in real life.

A picture model learns patterns from the pictures it receives. It has not touched fur or played with a dog. It often needs many examples. A model trained only on photos may struggle with a cartoon.

Recognize a question that does not fit

Imagine someone holds up a toaster and asks, "Cat or dog?" You can say, "Neither!" You understand that the question's two choices do not fit.

A simple model trained only on cats and dogs still chooses one label. It cannot add a new answer for itself. That makes your judgment important when its answer looks strange.

Different strengths

Machines can work through huge numbers of examples very quickly. They do not get tired or bored as people do. People bring meaning, feelings, and judgment to situations.

What you have seen matters too. A child who has met only brown dogs may be surprised by a white one. One new experience can change the child's idea. A simple model needs new examples and another training step to change its learned patterns.

AI can help with quick jobs that have clear patterns. It can turn speech into text or find beach photos. People need to lead jobs involving care, fairness, and responsibility.

Ask four questions

Before choosing AI, think about the job:

• Would AI help, or can I do this well myself?
• Could a mistake hurt someone or cause trouble?
• Does this need feelings, fairness, or a person's judgment?
• Would I need to share private information?

These questions help you choose a useful role for the tool.

Work through two choices

Your class needs a recycling-poster title. An approved AI tool could suggest ideas. You discuss the ideas and choose together. AI supplies possibilities; people judge which title fits.

Now your class must choose an art-prize winner. This affects people's feelings and needs a fair decision. People should discuss the work and make the choice. A computer's fast answer cannot take responsibility for that choice.

Keep the practice that helps you learn

If your task is to practice writing a story, having AI write it takes away your practice. Write the story yourself. A spelling check afterward may help if your teacher allows it.

Ask what skill the task is helping you build. Getting a finished answer and learning how to make one are different things.

People take responsibility

Important decisions about health, safety, places, and prizes need a person to check and decide. A person can explain a decision and take responsibility. An AI tool cannot do that.

There is also a cost to running AI. Large systems work in data centers, buildings full of computers. They use electricity and often water for cooling. Choosing AI when it really helps avoids using it for jobs you can easily do yourself.

A chatbot is a program you talk with by typing or speaking. An AI chatbot uses a language model, trained on huge amounts of writing. It builds an answer piece by piece by predicting words that fit learned patterns.

A chatbot is not reading one checked answer from a trusted book. It can produce fluent writing with invented facts. An image generator learns from pictures paired with descriptions. It creates a new picture using patterns connected with your words.

Give a clear prompt

A prompt is your question or instruction to an AI tool. The tool also uses its training, earlier chat messages, and instructions built in by its makers. The prompt is the part you can change.

Compare these two image prompts:

• "Draw a dog."
• "Draw a cartoon of a small brown dog with floppy ears catching a red ball in a sunny park."

The first leaves many choices open. The second names a style, size, color, action, and place. Each detail helps the tool make something closer to your idea.

Improve one result at a time

Read or look at the result. Decide what needs changing. Then change the prompt or ask for a clearer result.

Suppose a story is too long for a younger child. You might ask, "Make it shorter, about 5 sentences, for a 7-year-old." This tells the chatbot both the length and the reader. If one detail is wrong, name that detail in your next request.

Asking again is a normal part of creating with AI. It still does not guarantee that every detail will be right. You need to inspect each result.

Real-looking does not mean real

Generated pictures can show events that never happened. Chatbots can invent quotes and book titles. Treat these creations as things to examine before using them.

The same prompt may produce different answers on different tries. Generative AI creates from learned patterns each time; it does not simply fetch one fixed answer.

An AI answer is a starting point for checking. It may sound smooth and still contain errors. Exact names, dates, and numbers need close attention.

Follow a fact back to a trusted source

Suppose a chatbot says your town's mayor started in 2019. The town's official website says 2021. Use the official date and fix your work. Check the other facts in the chatbot's answer too.

Compare important facts with at least two trusted sources. An official website, library book, encyclopedia, or knowledgeable teacher can help. Read what each source actually says. Asking the same chatbot again does not give you an independent source.

Also ask whether the claim makes sense with what you know. A surprising detail is a reason to investigate carefully.

Look for unfair patterns

Facts are only one part of checking. Look at who appears in text or pictures. Does a picture of scientists show only one kind of person? Does an answer leave out people who should be included? The output may repeat a narrow pattern from training data.

Can you tell whether a picture is AI-made?

Some pictures have clues: jumbled writing, strange fingers, odd reflections, or objects melting together. These can make you question a picture's origin.

But AI pictures can also look smooth and ordinary. Missing visual clues cannot tell you that a picture is real. Check who shared it, whether it has an AI label, and whether trusted news sources report the event.

If you cannot tell whether something is real, keep it unshared while you check.

Does it suit your reader?

An answer can be factually correct yet unsuitable. A younger child may need simpler words. A poster may need a shorter explanation. You can edit the result yourself or ask for a clearer version. Then check the changed version too.

Start with the AI tools your school or family has chosen. Ask a parent, caregiver, or teacher before using one. Many chatbots have age rules; some require users to be at least 13. Your adult can check which tools are suitable for you.

Protect private information

Keep your full name, address, phone number, school name, and passwords out of AI chats. Do not upload photos of yourself or other people. A company may save the information you send.

You can describe an imaginary character without using a real child's details. A story prompt can say "a curious child" without naming your school or home.

Be honest about help

Follow your teacher's rules about AI. If AI helped with allowed work, say how. For example: "I used an AI tool to get ideas for my title."

Keep yourself as the main creator. Think, write, and draw for yourself. Make sure you understand the work you hand in. Copying a complete AI answer can leave you unable to explain it.

Think about the person who could be hurt

Never use AI to make fake pictures, voices, or messages about real people. Do not use it to trick someone or be mean. A fake about a classmate can hurt even if its maker calls it a joke.

Turn to people for care

A chatbot can sound friendly, but it is not a person or a friend. It has no feelings and cannot care for you as people do. Talk with a trusted adult about worries, strong feelings, or problems.

If a tool shows upsetting, scary, or rude content, stop and close it. Tell a trusted adult. You have done nothing wrong by asking for help.