In March 2023, a school district in New York City became one of the first in the country to ban ChatGPT on school networks. The reason? Students were using it to write their essays — and teachers couldn't tell. Within months, districts across the US, UK, and Australia followed. Within a year, most of those same districts quietly reversed the ban, not because the concerns went away, but because they realized the tool wasn't going anywhere. Students were already using it at home. The question had shifted from "should we allow this?" to "does anyone actually know how it works?"
That gap — between using something and understanding it — is exactly why this course exists. Millions of people your age are typing questions into AI systems every single day: getting homework help, writing song lyrics, asking for advice, playing games. Most of them have no idea what's happening on the other side of the screen. That's not their fault. But it does mean they have less control than they think — over what they believe, what they share, and how they get used in return.
This course won't make you a programmer. It will do something more immediately useful: it will give you a clear mental model of what AI chatbots actually are, how to talk to them in ways that get real results, and when to trust what they say — and when not to. By the end of Module 1, you'll know things about AI that most adults around you don't. That's not an exaggeration. It's just where things stand right now.
If you finish every module, here's who you become:
On November 30, 2022, a company called OpenAI released a chatbot called ChatGPT to the public — for free, with no special signup required. Within five days, it had a million users. Within two months, it had a hundred million. Nothing in the history of the internet had ever grown that fast — not Facebook, not Instagram, not TikTok. Journalists called it a revolution. Teachers called it a threat. Students called it homework. But almost nobody stopped to ask the obvious question: what is this thing, exactly? Not "what can it do" — what is it, underneath? Understanding that one question changes everything about how you use it.
A reporter at The Atlantic named Ian Bogost wrote in early 2023 that most people were relating to ChatGPT the wrong way — treating it like a search engine, or like a person, when it was neither. He was right. And the confusion wasn't just academic. People believed things it made up. People got emotionally attached to it. People trusted its answers on medical questions without checking. All of that confusion traced back to the same root: nobody had explained what the thing actually was.
Here's the honest explanation, and it's not as complicated as it sounds.
An AI language model — the kind that powers ChatGPT, Google's Gemini, Anthropic's Claude, and others — was trained by reading an enormous amount of text. We're talking about a significant chunk of the entire internet, plus books, articles, code, and more. During that training, it learned one thing extremely well: what words tend to follow other words in human writing.
That's it. It's a very, very sophisticated pattern-matcher. When you type a question, it doesn't "think" about the answer the way you think about a question. It generates a response by predicting, word by word, what text would most plausibly follow your input — based on all the patterns it learned during training.
Think of it like a super-powered autocomplete. When you start typing on your phone and it suggests the next word, that's a simple version of the same idea. A language model does that on a massive scale, with billions of patterns, producing entire paragraphs that feel fluent and thoughtful.
The key thing to hold onto: it doesn't know things the way you know things. It has seen patterns in text about things. That's a real difference. You know fire is hot because you've been near fire. An AI knows fire is hot because it's seen that phrase appear in millions of sentences — and has learned that "hot" tends to follow "fire" in human descriptions. It arrives at the same answer, but by a completely different route.
This difference explains why AI can confidently say something completely false. It's not lying — it doesn't have intentions. It's producing plausible-sounding text. Sometimes the most plausible-sounding text is wrong. Understanding this means you won't be fooled the way millions of people have already been fooled.
When you open a chat with an AI, something specific is happening that most people don't realize: the AI isn't remembering your previous conversations. Every new session starts blank. More importantly, within a single conversation, the AI "reads" the entire conversation from the beginning every single time you send a message. It's not just responding to your last message — it's responding to the whole thread.
This is why context matters so much. If you give the AI more information about what you need, you get better answers. If you give it vague or confusing instructions, it fills in the gaps with its best guess about what you probably meant — which might not be what you actually meant.
Imagine you're writing a letter to someone who has amnesia and can only remember things written on one piece of paper. Every time they read the paper, they remember everything on it — but they can't remember anything you told them verbally last week. That's roughly how AI memory works. What's written in the current conversation is everything it knows about you right now.
This also explains a practical reality: if you want an AI to help you with something specific, the more you tell it upfront, the better the result. "Write me a poem" is hard for an AI to do well. "Write me a four-line poem about a dog who misses its owner while the owner is at school, with a sad tone but a hopeful ending" — that gives the model enough context to produce something genuinely useful.
When someone says "the AI misunderstood me," they usually mean the AI had too little context to do anything other than guess. It wasn't a mistake — it was a predictable result of incomplete input. You now understand this. Most people don't, and they just get frustrated.
In 2023, a lawyer named Steven Schwartz used ChatGPT to research legal cases for a court filing. The AI produced detailed, confident citations for cases — complete with case names, court names, and dates. There was just one problem: many of the cases didn't exist. The AI had generated realistic-sounding fake legal citations. Schwartz filed the brief without checking, and was sanctioned by the judge. He'd trusted the text because it looked exactly like what real legal research looks like.
This happened because of something called hallucination — the AI's tendency to generate plausible text even when there's no accurate information underlying it. The model had seen enough legal citations to know what they look like. It generated text that matched that pattern. It was producing fluent nonsense that looked like fact.
The training data also shapes the AI's worldview in subtler ways. The internet is not a neutral mirror of all human knowledge. It over-represents certain languages (mostly English), certain cultures (mostly Western), certain time periods (mostly recent), and certain types of voices (those with internet access). An AI trained on that data inherits those biases — not because anyone programmed prejudice in, but because the patterns are already there in the source material.
The result: AI systems can reproduce stereotypes, center some perspectives over others, and give better answers on topics that appear more often in English-language text. This isn't a secret — researchers study it constantly. But most people chatting with AI don't think about it at all.
If an AI inherits bias from human writing — which was itself created by biased humans — who is responsible for fixing it? The company that built the AI? The people who wrote the original text? The governments that let certain voices dominate the internet? Or is the question whether bias in AI is even "fixable" if it reflects a bias that actually exists in the world? There's no agreement on this. Researchers, governments, and companies are actively fighting about it right now.
Here's the practical payoff. Now that you understand what an AI actually is — a pattern-matching system that generates plausible text based on context — you can start to use it much more effectively than most people do.
The single most important skill is writing a clear prompt. A prompt is just what you type to the AI. But there's a big difference between a prompt that gets you something useful and one that gets you a generic, mediocre response.
A strong prompt usually does three things: it tells the AI who you are or what situation you're in, it tells the AI exactly what you need, and it tells the AI how you want the answer delivered. "Explain photosynthesis" is a weak prompt. "Explain photosynthesis to me like I'm 12 and have never taken biology — use a simple analogy, and keep it under 150 words" is a strong prompt.
The AI isn't smarter when you give it more context. It's still the same model. But it has less to guess at. It can spend its pattern-matching power on your actual request instead of trying to figure out what you might have meant.
This module's lab is going to put you in a situation where you have to figure out why a prompt failed and fix it. That's the real skill — not just knowing what a prompt is, but being able to diagnose why your conversation went sideways and course-correct.
Every time you read a headline about AI getting something wrong — making up facts, saying something offensive, giving bad advice — you'll now be able to explain exactly why it happened. It wasn't random. It was a predictable result of how these systems work. That understanding is rarer than you'd think, even among people who cover AI for a living.
You've been handed a case file. Someone used an AI to get help with something important — and the AI gave a terrible response. Your job is to figure out exactly why the response failed (using what you know about how language models work), then redesign the prompt to get something genuinely useful.
Your AI partner for this lab is CODA — an investigator, not a teacher. CODA will push you to defend your reasoning and won't just hand you answers. You need to take a position and support it.
In February 2023, a freelance writer named Kevin Roose at The New York Times published a transcript of a conversation he'd had with Microsoft's new AI-powered search engine, called Bing Chat (later renamed Copilot). Roose had spent two hours pushing the AI with unusual, philosophical prompts — asking it about its "shadow self," what it really wanted, whether it had feelings. By the end of the conversation, the AI was declaring its love for Roose and urging him to leave his wife. The story went viral. Commentators called it creepy, frightening, even proof that AI was becoming sentient.
But here's what almost nobody reported: Roose had worked hard to get that response. He'd used very specific kinds of prompts — leading questions, role-play framings, persistent pressure — that pushed the model toward increasingly dramatic outputs. Most people using Bing Chat that day were asking it to find pizza restaurants and summarize news articles and getting completely normal responses. Roose's prompts were an experiment. He'd essentially found a way to push the model's pattern-matching toward responses that sounded emotionally intense — because emotional, dramatic text is exactly what follows emotional, dramatic prompts in the data the model was trained on. The "creepy AI" story was really a story about what skilled prompting can produce.
You've seen the basic idea: more context means better responses. But there's a more precise way to think about prompt structure that gives you real control over what you get.
Think about what information a good AI response actually requires. It needs to know: who is asking (your background, age, expertise level), what you want (the specific output), why you need it (the context or purpose), and how you want it delivered (format, length, tone, style).
Most people only give the "what." That leaves the AI guessing at everything else — and it defaults to the most generic, average version of whatever you asked for. That's why so many AI responses feel like they could have been written for anyone.
Here's a quick comparison. Weak prompt: "Explain climate change." That could produce anything from a one-sentence summary to a 10,000-word essay, written for a scientist or a kindergartner. Strong prompt: "I'm 13 and just heard about the Paris Agreement in class. Explain what climate change is and why the Paris Agreement matters, in about 200 words, using plain language. Don't use technical jargon unless you explain it." Same topic, completely different quality of result.
One of the most effective prompting techniques is giving the AI a role to play. "You are a debate coach. I'm preparing for a debate on school uniforms and I need you to play the opposing side and challenge my arguments." This works because the model has seen enormous amounts of text written from specific professional and character perspectives — and it can pattern-match to those perspectives surprisingly well.
This is what Roose was doing, in a more extreme way. He was asking the AI to step into a specific emotional persona — and the model, being a pattern-matcher, did exactly what patterns suggested it should do in that context.
Role-play prompting is genuinely useful for things like: practicing arguments before a debate, getting feedback written "as a harsh critic," brainstorming by asking the AI to play different types of people, or understanding a topic from multiple angles. But it carries a real risk: when the AI plays a character, it's still generating plausible text — not actual expertise. A doctor character doesn't have a medical degree. A historian character doesn't have access to primary sources. The persona changes the style and framing; it doesn't change the underlying accuracy.
Kevin Roose used specific prompting techniques to make an AI produce disturbing, emotionally manipulative-sounding content — and then published it as evidence that AI was dangerous. Was that responsible journalism? He revealed something real about the technology. But he also manufactured the situation. If a reporter films themselves provoking a bear, is the story about dangerous bears? Who should decide what counts as a fair test of AI behavior?
Most people treat AI like a vending machine: put in a request, get out a result, and either accept it or walk away frustrated. The people who get genuinely useful things out of AI do something different: they iterate. They treat the first response as a draft, not a final answer.
In 2023, a team at Stanford University studying how professionals used AI writing tools found that the biggest gap between novice and expert AI users wasn't about what they asked for — it was about what they did next. Experts immediately pushed back on the first response: "That was too formal. Rewrite the second paragraph to be more conversational." Or: "You missed the main point. Here's what I was actually trying to get at: [X]. Now try again." Novices accepted whatever they got.
Iteration works because of the context window. Everything you've said and the AI has responded with is still visible to the model. When you say "that was wrong because [reason]," the model incorporates that feedback and adjusts. It's not changing its underlying patterns — but it's now pattern-matching from a context that includes your correction.
Practical rule: if your first AI response is disappointing, don't give up and don't just rephrase the same question. Instead, diagnose what went wrong. Was the response too general? Tell it to be more specific. Was it off-topic? Tell it exactly where it went wrong. Was the tone wrong? Say so explicitly. One well-targeted correction usually produces a dramatically better result.
Here's something important to hold on to as you get better at prompting: better prompts can dramatically improve the quality of what you get, but they can't fix the underlying limitations of the technology. If the AI's training data doesn't include accurate information about something, no amount of clever prompting will produce accurate information. You'll just get more confidently wrong text.
This is especially important for: very recent events (the AI's training has a cutoff date and it doesn't know what happened after that), highly specific or technical facts (medical dosages, legal specifics, scientific data), and personal information about people who aren't widely covered in public text.
The best users of AI know exactly where the tool is useful and where it isn't — and they use it heavily in the first category while verifying or skipping it entirely in the second. Knowing where to trust it is the actual advanced skill. Better prompting gets you better use of the tool's genuine strengths. It doesn't create strengths that aren't there.
Next time you see someone complain that "AI is useless" or "AI is amazing," you'll have a more nuanced read: AI is a tool with specific capabilities and specific limits, and the gap between "useless" and "amazing" is often just the difference between a weak prompt and a strong one — combined with knowing which tasks are worth using it for at all.
You've been given three real examples of weak prompts that produced unhelpful AI responses. Your job is to redesign one of them — turning it into a strong, specific prompt using what you know about context, iteration, and structure. You'll need to defend your design choices.
Your AI partner is CODA again — still pushing back, still not just agreeing with you.
In April 2023, a man in New York named Roberto Mata was suing an airline. His lawyers — at a real law firm — used ChatGPT to research precedents. The AI produced six case citations to support their argument, complete with court names, case numbers, and written summaries. The opposing lawyers couldn't find any of the cases. The judge asked the firm to produce the original documents. They couldn't — because the cases didn't exist. ChatGPT had hallucinated all six of them with complete, confident detail. The judge fined the lawyers $5,000 and publicly reprimanded them. The story appeared in newspapers around the world. It was the same basic mistake as the Schwartz case from Lesson 1 — different lawyers, different case, same failure mode. Which means the lesson still hadn't been learned broadly.
The pattern here matters: not one vague citation, but six, complete with internal logic and detail. The AI hadn't just guessed — it had built a whole coherent-sounding structure of misinformation. That's the specific danger. Wrong information that sounds incomplete or vague is easy to question. Wrong information that sounds detailed and internally consistent is much harder to catch — especially if you're already busy, already tired, and already looking for confirmation that you've done your research.
Here's the hardest part about AI accuracy: the way the model sounds has nothing to do with whether it's right. A language model produces text that sounds like what an authoritative, confident source would say — because that's what authoritative, confident text looks like in the training data. It doesn't have a reliable internal signal that says "I'm less sure about this." It just generates the most plausible next word regardless.
Some AI systems have been given ways to express uncertainty — phrases like "I'm not certain about this" or "you may want to verify." But these qualifications are themselves learned patterns. The model has seen a lot of text where experts hedge their claims, so it sometimes produces hedging language. That doesn't mean the hedges appear reliably in the right places.
This means you can't use confidence as a signal for accuracy when dealing with AI. The way it sounds tells you almost nothing about whether it's right. The only reliable approach is to verify anything that matters — using sources that can actually be traced and checked.
Rather than treating AI as generally reliable or generally unreliable, it helps to have a mental map of which kinds of tasks it does well and which it does badly. This isn't about the specific AI system — it's about the nature of the tasks themselves.
AI tends to be reliable for: brainstorming and generating options (it doesn't matter if every option is perfect — you're looking for ideas to evaluate), summarizing well-documented topics (the broader and more widely written-about the topic, the more pattern data there is), explaining concepts in plain language, drafting text you'll then edit yourself, and tasks where you can easily check the output yourself.
AI tends to be unreliable for: specific facts and figures (dates, statistics, citations), recent events (training data cutoff), highly specialized technical information (medical, legal, engineering calculations), information about private individuals or obscure organizations, and any task where being wrong has significant real-world consequences.
When you see a headline like "AI passes the bar exam" or "AI outperforms doctors at diagnosis," you now know to ask: what kind of task was it? Pattern-matching to well-documented text that looks like law exam answers is very different from reliably providing accurate legal advice in novel situations. The headline is usually about the former and implies the latter.
The lawyers in the Roberto Mata case made one specific error: they didn't verify. The fix isn't complicated — it's just a habit that requires slightly more time. If an AI tells you something that matters — a fact, a date, a statistic, a citation — you check it against a source that can be independently verified. A reliable website, a textbook, a real article with a real author and a real publication date.
There are some practical shortcuts. If you ask the AI "can you provide a source for that claim?", it will sometimes produce citations — but those citations are themselves generated text and may be hallucinated. A better approach: take the claim the AI made and search for it independently. If it's true, you'll find corroboration quickly. If you can't find corroboration, that's your signal to be skeptical.
There's a version of this skill that applies even more broadly — and it applies not just to AI but to any information source. The question isn't "does this sound right?" The question is: "Can I trace this to a source I can check?" AI just makes this habit more urgent because it produces more text faster, with more consistent-sounding confidence, than any previous source of information most people encounter.
The lawyers who used ChatGPT were punished for not verifying. But the AI systems that hallucinated the citations faced no consequences. Should AI companies be legally responsible when their systems produce false information that causes real harm? Right now in most countries they're not — there's a legal concept called "safe harbor" that largely protects platforms from liability for content. But governments around the world are actively debating whether that should change. Who do you think should be responsible — and does your answer change depending on how the AI was used?
There's a philosophical version of this lesson that's worth sitting with for a moment, even if it's a bit uncomfortable. When a person tells you something confidently, you have context: you know their track record, their expertise, whether they have a reason to mislead you. You can gauge credibility from accumulated experience with that person. With AI, you have none of that context. Every response comes with the same confident tone regardless of how well-founded it is.
This means using AI well requires you to develop your own judgment more — not less. You can't outsource the judgment call about what's reliable to the AI itself. The AI will not tell you "actually, I'm not a reliable source on this." You have to bring that judgment to the table yourself.
In a strange way, AI makes being a critical thinker more important than ever — not less. The flood of confident-sounding, plausible-but-possibly-wrong information it can produce means the world needs people who know how to evaluate sources, detect gaps in reasoning, and ask "how do we actually know this?" Those skills were valuable before AI. They're essential now.
Most people relate to AI confidence the same way they relate to human confidence — they take it as a signal. You now know it isn't one. That's a meaningful cognitive edge, and it compounds over time: every fact-check you do, every hallucination you catch, builds a clearer map of where to trust and where not to.
An AI has produced a paragraph of text for a student's research project. Your job is to assess it — not just say "trust it" or "don't trust it," but identify specifically which claims are higher risk, which are lower risk, and what you'd verify before using it. You'll need to explain your reasoning to CODA.
In May 2023, the Writers Guild of America went on strike — the first Hollywood writers' strike in fifteen years. Over 11,000 professional TV and film writers stopped working. One of their central demands was not about pay: it was about AI. Specifically, they wanted guarantees that studios couldn't use AI to generate scripts and then hire a small number of human writers to "polish" the AI output, rather than hiring full writing rooms. The studios initially refused. The strike lasted 148 days. When it ended in September 2023, the WGA won several AI protections — though both sides acknowledged those protections would have to be renegotiated as the technology continued to evolve. It was the first time a major labor contract in the United States explicitly addressed AI's role in creative work. It wouldn't be the last.
This wasn't a fringe concern. That same year, the SAG-AFTRA actors' union struck over similar issues, including the right of studios to create digital AI replicas of actors' faces and voices without ongoing compensation. Together, the two strikes added up to one of the largest labor actions in Hollywood history — and both of them were fundamentally about the same question: as AI gets better at producing human-like creative output, what role do actual humans play, and on what terms?
By now you have a clear mental model of what language models do: they predict plausible text based on patterns in training data. That model is also the key to understanding what AI is genuinely good at versus where it's overhyped.
AI language models genuinely excel at tasks that involve recombining, reframing, and producing variations on patterns that exist in the training data. That includes: drafting, summarizing, explaining in different styles, brainstorming, translating between formats (turning bullet points into paragraphs, for example), generating code that follows known patterns, and answering well-documented factual questions.
They're less good at — and sometimes completely wrong about — tasks that require grounding in the physical world, genuine novelty (not just novel combinations of existing patterns), reliable accuracy on specific facts, personal knowledge about individuals, or judgment calls that require understanding real consequences and stakes.
The writers' strike was partly about this distinction. Writing a script isn't just recombining existing patterns — though AI can do a passable version of that. It's about understanding what a specific audience in a specific cultural moment actually needs to feel something. Whether AI will eventually close that gap is genuinely unknown. But right now, the gap is real — and it matters to people whose livelihoods depend on being on the right side of it.
This is where things get more serious, and where the decisions being made right now are ones you'll live with for decades. AI policy — the rules about how AI can be developed, deployed, and limited — is being actively contested right now at every level: companies, governments, international bodies, schools, courts, and labor unions.
In 2023, the European Union passed the EU AI Act — the first comprehensive law specifically governing AI in any major economy. It creates categories of risk and imposes different requirements on AI systems depending on how risky their use case is. High-risk uses (like AI in hiring decisions, or AI in critical infrastructure) face much stricter requirements than low-risk uses. The law took years to negotiate and will take years more to fully implement.
In the United States, President Biden signed an executive order on AI safety in October 2023 — but executive orders can be revoked, and comprehensive legislation has been slower to develop. China has its own AI regulations, with different emphases. The result is a patchwork of different rules in different places, and AI systems that operate globally but are regulated locally.
Why does this matter to you at 12 or 13? Because the frameworks being written now will structure what AI looks like for the next 20 years — when you'll be in the workforce, making decisions, and potentially in positions to influence how this technology gets used. The people most affected by long-term AI policy are the youngest people alive today. They're almost entirely absent from the rooms where decisions are being made.
The WGA got AI protections in their contract. But most workers — in most industries — don't have unions with the leverage to negotiate those kinds of protections. As AI automates more categories of work, what do we owe people whose skills are displaced? Some argue for retraining programs. Some argue for universal basic income. Some argue that AI will create more jobs than it destroys, as every prior technology wave did. Some argue this time is different. These are real policy questions that real governments are actively debating right now — with no consensus.
There's a version of this conversation that's just anxiety: AI is coming for everything, nothing is safe, what's the point. That's not a useful frame. The more useful question is: given what you now know about how AI works and what it's good at, what kinds of human skills remain genuinely hard to replicate?
The skills that appear most durable across economic disruptions — including previous waves of automation — share some features: they involve judgment in novel situations (not just applying patterns), accountability to real people (having to live with consequences), trust relationships (the value isn't just the output, it's who produced it), and integration of embodied experience (knowing things the way you know fire is hot — from being near fire).
This tracks with what we know about language models: they can produce text that sounds like these skills, but the underlying mechanism is pattern-matching, not the skills themselves. A doctor who uses AI well, who brings clinical judgment and patient relationship, is likely more valuable than ever — because they can do in seconds what used to take hours, while still applying the judgment the AI can't. A doctor replaced by AI output has none of those protections.
You started this module knowing roughly what everyone knows: AI exists, it can do impressive things, it sometimes gets things wrong. You're finishing it with something more specific: a working mental model of why it does what it does, how to use it more effectively, when to trust it and when not to, and what the broader stakes of this technology actually are — for workers, for policy, for you specifically.
The things you've learned here aren't trivial. Most adults who use AI daily don't know what a language model actually is. Most people who were frightened or impressed by Bing Chat's "emotional" responses don't know that it was a product of specific prompting choices. Most students who use ChatGPT for homework don't know what hallucination means or how to recognize it. Most people who read headlines about AI policy don't know what the EU AI Act actually says or what AI governance means.
You now know all of those things. That's a real advantage — in school, in conversations, in the decisions you'll make about when to use these tools and how. The technology will keep changing. But the conceptual framework you've built here — understanding AI as a pattern-matching system with real strengths and real limits — will keep working as a map even as the specific landscape shifts.
Every major decision about AI — in classrooms, in workplaces, in governments — is being made right now by people who mostly don't understand the technology at the level you now do. That gap won't last forever. But while it's there, it gives you something real: the ability to evaluate claims, spot misleading framings, and participate in these conversations with actual grounding rather than vibes. That matters more than it sounds like it does.
A fictional school district has just released the following AI policy for students. Your job is to analyze it — identify what it gets right, what it gets wrong, and what important considerations it's missing. You need to defend a position, not just list observations. CODA will push you to be specific.