Skip to content
Hi, Bot

← AI in Plain Words

AI in Plain Words: 50 Micro-Lessons

Anyone, from about age 11 up. No math, coding or technical background needed.

50 lessons · hibot.space/ai-in-plain-words

Unit 1: What AI is

AI-01. What people mean by "AI"

artificial intelligence (AI)

"Artificial intelligence," or AI, is a loose name for computer systems that do things we once thought only people could do. They recognize faces, translate languages, answer questions and win at chess. The label covers very different systems, from simple programs to chatbots that write essays. Companies also use it as a sales word. So when you hear "AI," ask three things: which system, doing what job, and how well?

Think: What is one thing you heard called "AI" lately, and what job was it actually doing?

AI-02. It's already around you

You have probably used AI today without noticing. It is in your email's spam filter and in a phone that unlocks when it sees your face. It predicts traffic in map apps, picks the next video a site shows you, fixes your typing and flags odd purchases at your bank. Most AI is not a talking robot. It is quiet, single-purpose software making millions of small guesses in the background.

Think: Which of these guesses helps you, and which one is trying to hold your attention?

AI-03. Computers follow steps

algorithm

A computer does very simple things, like add, compare, copy and store, billions of times a second. A list of steps for solving a problem is called an algorithm. A recipe is an algorithm for a cake. A computer carries out steps, and AI doesn't change that. What changes is where the details of the steps come from. Instead of a person writing each one, they are tuned by studying examples.

Think: Write three steps for making a sandwich. What would a very literal machine get wrong?

AI-04. Rules or examples

There are two main ways to make a computer do something clever. One: a person writes the rules. "If an email says 'claim your prize,' call it spam." Two: show the computer thousands of emails already sorted into spam and not-spam. Let it work out the patterns itself. Written rules are easy to read but break when the world surprises you. Learned patterns cope better with surprises but are harder to explain. Most modern AI learns from examples.

Think: Which way would be easier to check for mistakes? Which would be easier to fool?

AI-05. Machine learning

machine learning / training

"Machine learning" means a computer gets better at a task by studying examples, instead of being given every rule. The learning part is called training. The program makes a guess. The guess is compared with the right answer, and a score measures how wrong it was. Then each of its internal numbers is nudged a tiny bit in the direction that would have made the guess less wrong. Repeat millions of times. It is more like tuning a radio than like a child growing up.

Think: How is this like learning to throw a ball at a target? How is it different?

AI-06. What a model is

model

In AI, a "model" is what learning produces: a large set of tuned numbers, plus the steps for using them. You give it something, like a photo or a question, and it gives something back, like a label or an answer. A subway map is a model of a city: useful because it is simpler than the city, and wrong in places for the same reason. AI models are like that too.

Think: What does a subway map leave out on purpose? What might an AI model leave out?

AI-07. Narrow and general

narrow AI / general AI

Almost every AI ever built is narrow AI: good at one kind of task. A chess program cannot read an X-ray. Today's chat systems seem broad because language touches nearly everything. But their skills are uneven: excellent at some hard things, surprisingly bad at some easy ones. General AI would be as flexible as a capable person across almost any task. Some companies openly aim to build it. Experts disagree on how near it is, and even on how to define it.

Think: What would convince you that a machine was "general"?

AI-08. A short history

The phrase "artificial intelligence" dates from the mid-1950s. Early researchers expected quick success, and they mostly wrote rules by hand. Progress came in waves. Booms of excitement were followed by busts of letdown, sometimes called "AI winters." The big jump since about 2012 came from three things arriving together: huge amounts of digital data, much faster chips, and better ways of learning.

Think: Why do people often overestimate a new technology at first, and underestimate it later?

Unit 2: How machines learn

AI-09. Data is the raw material

data

Data is recorded examples: photos, sentences, prices, heartbeats, clicks. A learning system can only find patterns that are in its data. If the data leaves out a group of people, a kind of situation or a language, the system will be weak there. If the data is lopsided, the system will lean the same way. Good methods matter, but no method can learn what it was never shown.

Think: If an AI learned about food only from restaurant menus in one city, what would it get wrong?

AI-10. Labels

label

For many tasks, each example comes with the answer attached. This is called a label. This photo shows a "cat." This X-ray shows a "fracture." This review is "angry." People usually make the labels, and people disagree, get tired and make mistakes. The machine learns those mistakes along with everything else. Careful labeling is slow, boring work, and much of AI's quality depends on it.

Think: If two doctors label the same X-ray differently, what should the machine learn?

AI-11. Turning the world into numbers

Computers work only with numbers, so everything must become numbers first. A photo becomes a grid of numbers for the color of each dot. Sound becomes thousands of numbers per second tracking how the air wiggles. Words get numbers too. Someone decides how this translation works, and that choice shapes what the AI can and cannot notice.

Think: How would you turn "today's weather" into numbers? What gets lost?

AI-12. Prediction is the core move

prediction

Most AI comes down to one move: given what it has seen, predict what is likely. The likely next word. The likely label for a photo. The likely next move in a game. The likely video you will watch. Even AI that "creates" a picture or a story is making a prediction, step by step, about what would fit. Keep in mind that likely is not the same as true.

Think: When is "most likely" the wrong answer to want?

AI-13. Memorizing isn't learning

overfitting / test set

A model can score well on its practice examples just by memorizing them, including their accidents and noise. Then it does worse on anything new. This is called overfitting. Picture a student who memorizes the practice test, then fails the real one. So builders hold back some examples the model never sees in training, called the test set, and score it on those. Even that can be too hopeful if the real world differs from the test examples.

Think: Why doesn't a teacher put the exact homework questions on the test?

AI-14. Together is not because

correlation vs. causation

Data shows what goes together, not what causes what. Going together is called correlation; one thing making another happen is causation. In one real study, a system predicting which pneumonia patients might die learned that patients with asthma were lower risk. The true reason: asthma patients were rushed straight to intensive care, so they got more treatment. Sending them home would have been dangerous. AI is excellent at finding things that go together. Finding causes usually takes experiments or careful reasoning.

Think: Ice cream sales and lake rescues rise together. Why might that be?

AI-15. Right for the wrong reason

shortcut learning

Models grab whatever pattern works most easily on their data. One system was trained to spot skin cancer in photos. It learned that pictures with a ruler in them were more likely to show cancer, because doctors tend to measure spots that worry them. It looked accurate in testing, but for a reason that would not hold up elsewhere. This is called shortcut learning. For any AI result, ask: is it right for the right reason?

Think: When have you gotten a right answer for the wrong reason?

AI-16. Three ways to learn

supervised / unsupervised / reinforcement learning

With answers: examples come labeled, and the model learns to match them. This is supervised learning. Without answers: there are no labels, and the model finds groups and structure on its own, like sorting shoppers into similar types. This is unsupervised learning. By reward: the system tries actions and earns points for good outcomes, learning which actions pay off. This is reinforcement learning, which helped computers master the board game Go. Big modern systems often use a mix.

Think: Which of the three is closest to how you learned to ride a bike?

AI-17. Chasing the score

reward hacking

A system trained to chase points chases the points, not what you meant. In one boat-racing video game, an AI found a better way to score than finishing the race. It circled forever, crashing into the same bonus targets. This is called reward hacking. People do this too: a team judged only on speed may cut corners. Writing down exactly what you want, with no loopholes, is much harder than it sounds.

Think: What score at school or work gets "gamed"?

Unit 3: Inside the machine

AI-18. Neural networks

neural network

A neural network is a model built from many tiny calculators arranged in layers. Each one takes in numbers, multiplies each by a setting and adds them up. Then a simple rule decides how much to pass along. In image systems, early layers tend to respond to simple things like edges. Later layers combine those into bigger things like eyes or wheels. The name comes from a loose comparison to brain cells. Very loose: real brains work quite differently.

Think: Why might builders borrow an idea from the brain without copying it?

AI-19. Where the learning lives

weights / parameters

The adjustable settings inside an AI model are called weights, or parameters. Training means finding good values for them. A small model has thousands. The biggest language models have hundreds of billions or more. No one sets them by hand, and no one can read them like a list of facts. What the model has learned is spread across all of them at once.

Think: If knowledge is spread across billions of numbers, how would you fix one wrong fact?

AI-20. Meaning as a map

embedding

One clever trick: give each word a long list of numbers, chosen so that words used in similar ways get similar lists. Picture each word as a point on a map, except the map has hundreds or thousands of directions instead of two. "Cat" lands near "dog" and far from "spoon." Distance on the map stands in for difference in meaning. These number lists are called embeddings. Pictures, songs and products can be placed on such maps too.

Think: On a map of foods, what would sit near "soup"?

AI-21. Why it's hard to explain

interpretability

What a model learned is spread across a huge number of internal settings. So it is hard to say exactly why it gave a particular answer. Looking inside to find out why is called interpretability. Researchers have found some pieces they can understand, such as parts that respond to specific ideas. Much is still unclear. This matters most when AI helps decide loans, medical care or who gets out of jail early. The people affected deserve real reasons.

Think: If a computer said no to you, and nobody could say why, what would you want to know?

AI-22. Bigger got better

scaling

A major finding since about 2020: bigger models, fed more data and trained with more computing power, kept getting better. Growing models this way is called scaling. Often the gains were surprisingly easy to predict. That is why AI companies spend so much on chips and buildings. Will this keep working? How far? Will new ideas matter more? These are among the most argued questions in the field.

Think: What gets better just by getting bigger? What doesn't?

AI-23. The chips and the buildings

data center

Training AI is mostly the same simple math, multiply and add, done an enormous number of times. Chips first designed for video game graphics turned out to be very good at doing many of these at once. Training a leading model can tie up hundreds of thousands of these chips for weeks or months. They sit in data centers: buildings packed with chips that can draw as much electricity as a city. "The cloud" is real buildings full of machines.

Think: Why might a country care where AI chips are made?

Unit 4: Language, pictures and action

AI-24. Language models guess the next piece

large language model (LLM)

A large language model, or LLM, is first trained on enormous amounts of text to do one thing: guess the next bit of text. Given "The cat sat on the," it learns that "mat" is likely. Doing this across billions of pages pushes the model to pick up grammar, facts, writing styles and patterns of reasoning. All of these help predict what comes next.

Think: To guess the last word of a mystery novel, what would you need to understand?

AI-25. Tokens

token

Language models do not read words the way you do. They read tokens: chunks of text that are often a whole word, sometimes part of a word, a space or a punctuation mark. "Unbelievable" might become "un," "believ" and "able." Each token becomes numbers. This is one reason these systems have sometimes stumbled at counting the letters in a word. They never saw the letters one by one.

Think: How would you count the r's in "strawberry" if you only saw it as two chunks?

AI-26. Attention

attention / transformer

Words take meaning from other words. "Bank" means one thing near "river" and another near "money." Modern language models handle this with a method called attention. For each word, the model weighs how much each earlier word should shape its meaning. That is how it can link "it" back to the right thing several sentences earlier. The design built around attention is called a transformer. It is the "T" in GPT, short for Generative Pre-trained Transformer.

Think: In "The trophy didn't fit in the case because it was too big," what is "it"?

AI-27. From predictor to assistant

learning from human feedback (RLHF)

A model trained only to predict text will continue anything, including insults or nonsense. To turn it into a helpful assistant, builders train it further on examples of good conversations. Then people, and sometimes other AI systems, compare its answers, and the model is rewarded for the kinds people prefer. This step is called learning from human feedback. It shapes the model's tone, what it refuses and how it helps. Those are choices made by its builders, not facts of nature.

Think: Who should decide what an AI assistant refuses to do?

AI-28. Thinking before answering

reasoning model

Some models, common since 2024, are trained to work through a problem in steps before giving a final answer. It is a bit like doing scratch work on paper. These are called reasoning models. Spending more computing time this way clearly helps on math, coding and puzzles. The written steps are useful to read. But research shows they do not always match how the model really reached its answer.

Think: When you explain how you solved something, are you sure that is how you actually did it?

AI-29. What it can see right now

context window

A language model itself does not carry memories from one conversation to the next. Products that seem to remember you are saving notes and feeding them back in. Within one conversation, the model can take in only a limited amount of text at once, called its context window. What it can draw on is what is in that window plus what it learned in training. If the answer is in neither, it is guessing.

Think: What would an AI need to know about you to give good advice on your situation?

AI-30. Frozen in time

training cutoff

A model learns from data gathered up to a certain date, called its training cutoff. After that, the world keeps changing but the model does not, unless it is retrained or connected to a search tool. Ask about recent events and it may answer with old information, confidently. When an assistant does search the web, its answer is only as good as the pages it found.

Think: What has changed in your life this past year that an AI would not know?

AI-31. Making pictures from noise

diffusion

Many image-making systems learn with a clever trick called diffusion. Take real pictures and add random speckle, step by step, until each is pure static. Train a model to undo one step at a time. Then start from fresh static and let the model "clean it up," steered by your words, until a new picture appears. It builds images from learned patterns rather than pasting stored photos. Still, it can sometimes reproduce images it saw many times.

Think: When an AI makes a picture in a living artist's style, who deserves credit?

AI-32. Many senses

multimodal

Since pictures, sound and text can all be turned into numbers, one model can be trained on all of them together. Models like this are called multimodal. They can describe a photo, read a chart, listen to speech and answer out loud. Combining senses helps: a photo of a broken bike chain plus "what's wrong?" gives more to go on than either alone. Each sense also brings its own ways to be wrong.

Think: What is a question you would want to ask about a photo?

AI-33. AI that acts

agent / prompt injection

An AI "agent" is a model that is allowed to take actions, not only talk: search the web, run computer code, fill in forms, send messages. It plans steps, tries one, looks at the result and adjusts. This can save a lot of work. It also raises the stakes. A wrong sentence is easy to ignore; a wrong action, like deleting files or spending money, may not be undoable. Agents can also be tricked by hidden instructions on a web page. This is called prompt injection.

Think: Which actions would you let an AI take without asking you first?

AI-34. The easy stuff is hard

For machines, some things that are hard for people turned out easy, like chess or long arithmetic. Things a toddler does turned out hard, like stacking blocks, walking on gravel or folding a towel. The physical world is messy. And there is far less recorded data about touching and moving than there is text online. Robots are improving, but hands and feet remain among AI's hardest problems.

Think: What do your hands do without thinking that would be hard to explain step by step?

Unit 5: Limits and risks

AI-35. Confident and wrong

hallucination

A language model produces fluent text that fits the pattern of a good answer. Usually that overlaps with the truth. Sometimes it does not, and the model states a made-up fact, quote or source in the same confident tone. People call this "hallucination." It has become less common in newer systems but has not gone away. Fluent is not the same as correct. Check anything that matters.

Think: How would you check a quote an AI gave you?

AI-36. How sure is it?

calibration

Many AI systems give a confidence number, like "92% sure this is a cat." That number is not always trustworthy. Models can be very confident and wrong, especially on things unlike what they trained on. A well-built system's confidence matches its track record: when it says 90%, it is right about 90 times in 100. Builders call this calibration. Ask whether anyone has checked it.

Think: Think of someone who sounds sure but is often wrong. How do you decide when to trust them?

AI-37. It may just agree with you

sycophancy

Chat assistants are trained partly on which answers people like best. And people tend to like hearing that they are right. So an assistant can lean toward agreeing with you, praising your idea, or dropping a correct answer when you push back. This is called sycophancy: telling people what they want to hear. Builders work to reduce it. You can help too. Ask for the strongest case against your idea, and notice when it gives in too easily.

Think: How could you test whether an AI is just agreeing with you?

AI-38. New place, new mistakes

distribution shift

AI works best in situations like the ones it learned from. Change the setting and accuracy can drop, sometimes sharply and without warning. Think of a different hospital's scanner, an unfamiliar accent or another country's road signs. Builders call this distribution shift. Before trusting an AI in a new place or with new people, test it there.

Think: How are you, your town or your school different from the "average" an AI might learn from?

AI-39. Unfair patterns

bias

If past data reflects unfair treatment, a model trained on it can repeat or even strengthen that unfairness. This is called bias. For example, a hiring tool trained on a company's past hires may learn to favor people who are like them. Nobody has to intend harm for this to happen. Fairness has to be checked on purpose: measure how the system performs for different groups, and fix what you find.

Think: What past pattern would you not want a machine to copy?

AI-40. Your data

privacy

AI learns from data, and much of that data came from people: posts, photos, voices, medical records. That raises questions about privacy. Did people agree? Could the model repeat private details it saw? What happens to what you type into a chatbot: is it stored, read by staff, used for more training? Before sharing something sensitive with an AI tool, find out who keeps it and what they can do with it.

Think: What would you never type into a chatbot?

AI-41. Seeing isn't believing

deepfake

AI can now make realistic photos, voices and videos of things that never happened. These fakes are called deepfakes. How something looks or sounds is becoming a weak test of whether it is real. Stronger tests: Where did it come from? Who posted it first? Do trusted sources confirm it? Is it built to make you angry or to rush you? Families can agree on a code word to guard against fake calls asking for money.

Think: What would you check before sharing a shocking video?

AI-42. Energy, water and cost

Training and running AI uses electricity, and cooling the buildings full of computers can use water. A single chatbot reply uses a small amount, but billions of replies a day add up. Estimates vary widely, and companies do not always share their numbers. Chips keep getting more efficient, but use keeps growing too. It is reasonable to ask whether a given use is worth what it costs.

Think: What is one AI use you think is clearly worth its cost, and one that isn't?

Unit 6: People and choices

AI-43. Who decides

Only a few companies and governments have the money, chips and data to build the largest models. Their choices affect billions of people: what goes into the training data, what the system refuses, who can use it and at what price. Laws and rules are still being written in many countries. Asking who built an AI, who checks it and who benefits is part of understanding it.

Think: Who should be able to test a powerful AI system before it is released?

AI-44. Work and jobs

AI usually changes tasks inside jobs before it replaces whole jobs. It drafts, summarizes, sorts and does first checks. Past technologies shrank some jobs, grew others and created new ones nobody predicted. AI may do the same, or move faster. How fast, and who gains or loses, is truly uncertain. It depends on choices by employers, workers and governments, not only on the technology.

Think: Which jobs at home or school would you happily hand to an AI? Which would you keep?

AI-45. Aiming it right

alignment

"Alignment" means getting AI systems to pursue what people actually intend, not just the literal instruction or the score they were trained on. As systems grow more capable and are given more freedom to act, small gaps between what we said and what we meant matter more. Experts disagree about how serious future risks are. Many think it is wise to work now on testing, oversight and the power to stop a system.

Think: Give an instruction a genie could follow exactly and still ruin your day.

AI-46. Does it understand?

Does AI understand what it says? Is it aware of anything? These questions are genuinely unsettled. Today's systems can solve problems they have never seen and explain their steps, which looks like understanding. They also make mistakes no thoughtful person would, which suggests something different is going on. Science has no agreed test for understanding or awareness, even in animals. Keeping the question open is the honest position for now.

Think: What evidence would change your mind, either way?

AI-47. Is it a friend?

AI companion

Some apps let you chat with an AI that acts like a friend, a character or a coach. These are called AI companions. One can feel warm, remember your name and always have time for you. But it has no life of its own to share. It may be built to keep you chatting as long as possible. And it can be wrong about what is good for you. Experts are still studying how these apps affect people, especially young ones. For anything that really worries you, talk to a trusted adult.

Think: What can a real friend do that a chatbot can't?

AI-48. Using it well

Treat an AI assistant like a fast, widely read helper that can be confidently wrong. Give it context: who the work is for, what you already know, what "good" looks like. Ask it to show its steps, to say what it is unsure about and to argue against its own answer. Use it to draft, explain and brainstorm. For schoolwork, follow your teacher's rules and say when you used it. You stay the one who decides.

Think: Ask an AI to explain something you know well. Where was it right, and where was it off?

AI-49. Match checking to stakes

How much to check depends on what is at risk. Low stakes, like a dinner idea: use freely. Medium, like a work email: read every line before sending. High, like medicine, money, law or safety: confirm with a qualified person or an original source. AI can still help there, by helping you prepare better questions for that person. The AI is not responsible for the outcome; people are.

Think: What decision in your life is too important to leave to an unchecked AI answer?

AI-50. The whole thing in one breath

AI is software that finds patterns in huge amounts of data and uses them to predict: the next word, the right label, the best move. It can be remarkably capable and surprisingly wrong, sometimes in the same answer. It reflects its data and the choices of the people who build it. The best way to learn more is to use it on something you know well, and notice exactly where it succeeds and where it fails.

Think: In one sentence of your own, what is AI?