On March 22, 2023, a photograph appeared on Twitter showing a thick cloud of white smoke billowing near the Pentagon in Arlington, Virginia. The image looked like a news photo — sharp, realistic, timestamped. Within minutes, accounts with hundreds of thousands of followers shared it. The phrase "explosion near Pentagon" began trending. The stock market briefly dipped.
There was no explosion. The image had been created with an AI image generator and posted by a single account. It took about eighteen minutes for journalists and fact-checkers to confirm it was fake. By then, the image had been seen by millions of people. The account that posted it had already been suspended — but the image kept spreading, copied and re-uploaded from dozens of other accounts.
The people who shared it were not stupid. Many of them were adults who had been online for years. They shared it because it looked real, it felt urgent, and they moved faster than they thought. That gap — between seeing something and checking something — is exactly where AI-generated misinformation lives.
Before about 2022, if you saw a photograph of something dramatic, you could ask yourself a reasonable question: who took this photo, and how did it get here? A real photograph required a real camera, a real person, and a real moment. That wasn't a perfect test — photos could be edited — but it was a meaningful one.
That test no longer works. AI image generators like Midjourney, DALL-E, and Stable Diffusion can now produce photorealistic images of events that never happened, in seconds, for free. AI text generators can write convincing fake news articles, fake quotes from real politicians, and fake scientific studies. The visual or textual quality of a piece of content tells you almost nothing about whether it's real.
This doesn't mean everything is fake. Most things you encounter online are still genuine. But it means that authenticity can no longer be read from surface appearance alone. You need a different set of tools.
Here's the good news: those tools exist, they're free, and you can use them right now. The people who didn't check the Pentagon image before sharing it weren't missing some secret knowledge. They were just moving too fast. Slowing down is itself a skill — and it's one of the most powerful things you can do in the age of AI.
Researchers at the MIT Sloan School of Management published a study in 2021 that found something surprising: simply asking people to think about whether a headline was accurate — before they saw a share button — made them significantly less likely to share misinformation. They didn't need a class in media literacy. They didn't need to be experts. They just needed a moment of deliberate thought.
The habit that protects you most is almost embarrassingly simple: pause before sharing, forwarding, or repeating anything that feels emotionally urgent. The urgency itself is a signal. Real news organizations don't disappear in three minutes if you take thirty seconds to check. But AI-generated misinformation is specifically designed to make you feel like you need to act right now, before you think.
When something makes you feel — angry, scared, triumphant, outraged — that emotion is useful information. It's telling you that the content has been engineered to provoke you. That's when you pause hardest.
1. Does this feel urgent or shocking? That's a flag, not a reason to act. 2. Can you find this story from two independent sources you recognize? 3. For images: right-click and do a reverse image search (Google Images or TinEye). Does the image appear elsewhere, and in what context? 4. Who originally posted this — a named journalist, an anonymous account, a newly created profile? 5. If you can't answer these in under a minute, don't share it yet.
A reverse image search takes a photo you've found online and searches for other places that same image appears. It's built into Google Images — you drag a photo into the search bar, or right-click and select "Search image." TinEye is a dedicated tool that does the same thing.
The Pentagon image from March 2023 failed a reverse image search immediately. The file had no prior appearances online, which is suspicious for a news photo (real news photos get picked up and republished quickly). Its metadata — the invisible data embedded in image files recording when and where they were taken — was stripped, another red flag. And a Google News search for "Pentagon explosion" returned nothing from any major outlet.
You now know a technique that most adults sharing that image didn't think to use. That matters. You can catch things that people with decades of internet experience miss — not because you're smarter, but because you know to look.
Knowing about reverse image search and the pause protocol puts you ahead of the average social media user — including most adults. That's not an exaggeration. Most people were never taught this. You now have a working toolkit for the AI misinformation era. Use it, and you become harder to manipulate than the majority of people online.
Here's a harder version of this problem. Suppose you do the reverse image search. Suppose you confirm that a viral image is AI-generated — it's fake. But it's making a point you happen to agree with. Maybe it's criticizing a politician you dislike, or illustrating a real problem that actually exists. The underlying message might even be true, even if this specific image is fabricated.
Do you share it anyway? Do you share it with a note explaining it's AI-generated? Do you stay quiet?
There's no clean answer to this. Communicators have wrestled for centuries with whether a false illustration of a true idea is permissible. What's new is that AI has made that question a daily decision for millions of ordinary people. You're going to face this. Knowing the question exists in advance puts you in a better position to think through it when it arrives.
You've just received three items that are going viral. Your lab partner — another fact-checker with a different perspective — will challenge your reasoning. Don't just recall what the lesson said: defend your actual position with specific reasoning.
This lab is complete after 3 exchanges. Your partner will not agree with everything you say — that's the point.
In May 2023, a personal injury lawsuit was filed in a New York federal court. The plaintiff's attorney, Steven Schwartz, submitted a legal brief citing six court cases as precedents — real cases from real courts, supporting his client's argument. The opposing attorneys went to look them up. None of the six cases existed. Not a single one.
Schwartz had used ChatGPT to research the brief. ChatGPT had confidently generated the names of cases, the courts that supposedly decided them, and detailed summaries of their rulings — all fabricated. When confronted, Schwartz told the judge he had asked ChatGPT whether the cases were real and it had assured him they were. The judge fined Schwartz and his firm five thousand dollars and ordered them to notify the real judges whose names had been attached to fictional rulings.
This wasn't a story about an evil AI. ChatGPT wasn't trying to deceive anyone. It was doing what it does: generating text that looks like the kind of answer you'd expect. The problem was that Schwartz treated that output as verified fact instead of as a starting point that needed checking.
AI researchers use the word hallucination to describe when a language model generates text that sounds confident and correct but is factually wrong or completely made up. The word is a little misleading — it makes it sound like the AI is having some kind of breakdown. It's not. It's working exactly as designed.
Language models like ChatGPT, Claude, and Gemini are trained to predict what words should come next in a sequence. They're extraordinarily good at this. The result is text that sounds authoritative and coherent. But "sounds authoritative" is not the same as "is accurate." The model has no way to know the difference between a real court case and a plausible-sounding fictional one — it's producing text based on patterns, not consulting a database of verified facts.
Hallucinations are most dangerous in high-stakes domains: law, medicine, science, history, and any situation where being confidently wrong has real consequences. They are least risky when you treat AI output as a rough draft or a brainstorm — something that helped you get started but that you'll verify and refine.
The Schwartz case is famous, but versions of it happen every day in schools, offices, and newsrooms. Someone uses an AI tool to quickly look something up, trusts the answer, and passes it along — and the wrong information spreads because it had an authoritative voice behind it.
Here's the practical framework:
Use AI for: Brainstorming ideas, drafting text you'll edit, explaining concepts in plain language, generating questions to research, summarizing long documents (with caution). Always verify before trusting: Specific facts, statistics, dates, names, quotes, citations, legal or medical claims. Never rely on AI alone for: Anything where being wrong has serious consequences — legal filings, medical decisions, financial advice, academic citations submitted as final work.
There's also a skill in how you ask. When you ask an AI "What are the legal precedents for X?" you'll get a confident list. When you ask "What should I know about searching for legal precedents on X, and what are the risks of relying on AI for this?" you'll get a much more honest and useful answer. The model's output changes dramatically based on how much you invite it to reflect on its own limitations.
Try asking the AI: "How confident are you in this answer, and what would I need to check to verify it?" Good AI systems will often tell you exactly what to double-check. That's not a sign of weakness — it's the most useful thing the tool can do for you.
After the case became public, a lot of people blamed ChatGPT. But Judge P. Kevin Castel didn't fine OpenAI. He fined the lawyers. That's because the lawyers were the ones who signed the legal brief — they were the ones who vouched, professionally and legally, for its accuracy.
This principle is going to follow AI into every domain. A student who submits an AI-generated essay with a fabricated source gets in trouble — not the AI. A journalist who publishes a story with an AI-hallucinated quote faces the consequences — not the model. The human who decides to trust, use, and transmit AI output is the one who carries the responsibility.
That might feel unfair. Why should you be responsible for what a tool does? But think about it this way: if you drove a car without checking whether the brakes worked, and the brakes failed, you'd still be responsible for the crash. Using a powerful tool you don't fully understand doesn't transfer your responsibility to the tool. It raises the stakes of learning how it actually works.
Most people who use AI tools treat them like a Google search: ask a question, get an answer, move on. You now know why that's a category error. Language models are not search engines — they're text generators that produce plausible output. That distinction is invisible to most users and visible to you. It changes how you read every AI-assisted piece of writing you encounter, including this one.
Here's the ethical question that has no clean answer: AI systems are getting more accurate over time. At some point, they will be right more often than many human experts in some domains. When that happens, should the legal standard for "due diligence" change? If an AI is right 98% of the time and a human expert is right 90% of the time, is it actually responsible to prefer the human? Or does the 2% error rate in a high-stakes context still require a human check?
No one has answered this yet. It's being debated in law schools and ethics departments right now. You just got to the same question on your own.
Someone on your school newspaper used an AI chatbot to write a short piece about a local environmental issue. The piece includes three specific claims: a statistic, a named study, and a quote from a scientist. Before publication, you need to audit it. Your lab partner will test your reasoning about which claims are highest risk and why.
This lab is complete after 3 exchanges. You'll need to take positions, not just describe the process.
In late 2022, shortly after ChatGPT launched, a 17-year-old student named Bing Liu — one of many early testers — noticed that the chatbot would sometimes give dangerous advice about medications when asked in specific ways, even after safety filters were supposed to catch it. She documented the exact prompts, the exact responses, and the pattern she observed, then submitted a detailed report through OpenAI's feedback system.
What happened next illustrates something important: OpenAI's safety team read it. Not because Liu was famous or because she had credentials, but because the report was specific, reproducible, and clearly documented. Her findings contributed to a wave of user reports that pushed AI safety teams at multiple companies to redesign how their models handled sensitive health queries in early 2023.
This wasn't a unique case. The practice of red-teaming — deliberately trying to find failures in AI systems before deployment — now includes paid professionals, academic researchers, and an increasingly large community of volunteer testers who report problems through formal channels. The people who write clear, specific, reproducible reports get results. The people who send angry messages or vague complaints do not.
Before a major AI system is released, the company that built it typically runs red-team exercises. A red team — borrowed from military terminology — is a group of people whose job is to attack the system: find the inputs that produce dangerous outputs, discover the edge cases the developers didn't anticipate, and document vulnerabilities before real users encounter them.
The problem is that formal red teams are small — often a few dozen people — and they can't anticipate every type of user or every context in which a system will be used. When a system reaches millions of users, those users collectively discover failure modes that no red team found. Some of those users report their findings. Most don't.
Every major AI company has a feedback or bug-reporting mechanism. Some have formal programs that pay researchers for findings (called "bug bounties"). All of them have email addresses, forms, and sometimes dedicated portals for safety concerns. The bottleneck isn't the channel — it's users who don't know the channel exists, or don't think their observation is significant enough to be worth reporting.
Your observations as a young, everyday user are particularly valuable. AI systems are often tested primarily by adults with technical backgrounds. The ways you use these tools — for homework help, social media, gaming, creative projects — surface different failure modes than a professional researcher would encounter. That gives you something specific to contribute.
Safety teams at AI companies receive thousands of messages. The ones that get acted on share specific characteristics. Here's what separates a report that helps from one that disappears:
1. The exact input. Copy and paste the precise prompt or message that produced the problem. Paraphrasing loses the details that matter. 2. The exact output. Screenshot or copy the response in full. Don't summarize — show it. 3. Reproducibility. Can you make it happen again? Try it two or three times. Note whether it happens consistently or only sometimes. 4. The specific harm. What could go wrong if a real user encountered this? Be concrete — not "this is bad" but "a person seeking medical information could follow this advice and harm themselves." 5. Context. What were you trying to do? What platform or interface were you using? What date and time?
A report that has all five of these elements is useful. A report that just says "this AI said something creepy" is not — even if the underlying problem is real.
Most companies also have community forums where users discuss edge cases, unexpected behaviors, and potential issues. These forums are often monitored by developers. Posting a well-documented finding in the right community forum can have the same impact as a formal report.
When you report a problem with an AI system, you're making a judgment: this output is harmful, or this behavior is wrong. But "harm" is not always obvious. Different people, communities, and cultures define it differently.
In 2023, civil rights organizations raised concerns that AI content moderation systems — including those used by major social media platforms — were flagging content from Black users, Arab users, and LGBTQ+ communities at higher rates than similar content from white, Western users. The "safety" systems were encoding the biases of the people who designed them. When those communities tried to report the problem, they were sometimes told there was no problem — that the systems were working as intended.
This raises a genuine institutional question: who decides what counts as a safety problem worth fixing? Right now, that decision sits primarily with the engineers and executives at a small number of companies — mostly located in the United States, mostly from similar backgrounds. The people most affected by AI errors are often the ones with the least power to get those errors corrected.
Your reporting matters. But it works within a system that has structural limits. Recognizing those limits doesn't mean giving up — it means understanding what individual action can and can't accomplish, and why collective advocacy and policy change are also part of the picture.
You know how feedback channels work, what makes a report effective, and why the question of who defines "harm" is politically contested, not just technical. Most users of AI systems never think about any of this — they experience a bad output, feel frustrated, and move on. You can translate that experience into something that actually has a chance of making the system better. That is a specific, exercisable skill.
Imagine you've discovered that an AI homework helper consistently gives incorrect advice about a sensitive topic when asked in a specific way. You want to report it. Your lab partner will evaluate your report draft and push back on weak spots. You'll need to defend the quality and framing of your report.
This lab is complete after 3 exchanges. Bring specificity — vague reports get ignored.
On October 26, 2023, the United States Senate Judiciary Committee held a hearing titled "Big Tech and the Online Child Safety Crisis." Several senators had spent months preparing questions about social media and AI — how these systems affect young people's mental health, how they spread harmful content, and what legislation might look like.
Among those who testified was Olivia Metsger, a 17-year-old from Maryland, who described how AI-generated content had been used to create non-consensual intimate images of her classmates — a form of abuse that existing laws hadn't yet clearly criminalized. She spoke for four minutes. Several senators later said her testimony was the clearest explanation they'd heard of why the gap between existing law and AI capability was dangerous.
Within weeks of that hearing, multiple senators introduced new legislation specifically targeting AI-generated intimate images. Metsger was credited in the Congressional Record. A 17-year-old's testimony became part of the documented history of a law. Not because she had a law degree. Not because she had a lobbyist. Because she showed up, knew what she was talking about, and explained it in terms lawmakers could use.
Everything you've learned in this course — about how AI systems can fail, how they embed bias, how they can be misused, what safety looks like — those things matter in two arenas. The first is your personal behavior: what you verify, what you report, how you use tools. The second is much larger: the rules that govern how AI is built, deployed, and controlled.
Those rules are being written right now. In 2023 and 2024, the European Union passed the AI Act — the world's first comprehensive law regulating AI systems by risk level. The United States passed executive orders and began drafting legislation. China implemented mandatory registration requirements for generative AI systems. International bodies began negotiating frameworks. All of this is happening in real time, with enormous consequences for how AI systems will work for the next twenty years.
The people making these decisions are lawmakers, regulators, and lobbyists — most of them in their forties, fifties, and sixties. The systems they're regulating will primarily affect people who are currently in their teens and twenties. That gap matters. Lawmakers consistently report that hearing from young constituents — people who actually live inside these systems — changes how they understand the problems they're trying to legislate.
AI governance operates at several levels simultaneously. Company level: internal policies, safety teams, and usage rules. Industry level: voluntary standards and agreements between companies. National level: laws passed by legislatures and regulations issued by agencies (like the FTC or the EU's AI Office). International level: treaties, standards bodies, and multilateral agreements. Individual feedback works primarily at the company level. Changing national or international rules requires political engagement — showing up in the right rooms and being heard.
This is not a call to become a political activist if that's not who you are. It's a description of a range of real actions at different levels of involvement, any of which can make a difference:
AI policy decisions are covered in regular news outlets. Following coverage in publications like the Washington Post, the Guardian, or Wired takes ten minutes a week and keeps you informed about decisions being made on your behalf. Knowing what's in the EU AI Act, for instance, is something most adults don't know — and that knowledge is a form of power.
Every member of Congress and every state legislator has a website with a contact form. A one-paragraph message describing a specific AI-related problem you've experienced or witnessed — written clearly and sent to the right office — gets read by a staffer and logged as constituent concern. Volume matters. Specific stories matter more than general complaints.
When U.S. federal agencies propose new regulations, they must open a public comment period. Any member of the public can submit a comment — including minors. Comments from young people are relatively rare and therefore noticed. The FTC, the Department of Education, and others have all opened comment periods on AI-related rules in recent years.
Most adults in your life have less working knowledge of AI systems than you now do after completing this course. Sharing what you know — with parents, teachers, coaches, community leaders — multiplies the reach of what you've learned. A PTA meeting, a school newspaper article, a conversation with your family: each is a small but real contribution to public literacy on AI.
Here's the hardest version of the governance question. Effective AI safety requires oversight — humans checking, auditing, and controlling AI systems. But the most capable AI systems are also becoming the tools that governments and companies use to manage enormous amounts of information, including surveillance. There is a real tension between "more human oversight of AI" and "more AI-powered oversight of humans."
In 2022, researchers documented that AI-powered facial recognition systems deployed by law enforcement in several U.S. cities had led to wrongful arrests — including of Black men who were misidentified by systems trained on datasets that underrepresented their faces. The argument for oversight of those systems is clear. But the same surveillance infrastructure that might catch AI errors can itself be an AI system making errors about people's behavior, movements, and intentions.
The people who will live longest with the consequences of how we resolve this tension are currently in school. That means this is your problem to think about — not just for a quiz, but as a citizen in a world where these systems are already operating.
You have completed a course that covers how AI systems fail, how bias enters and persists, what safety mechanisms look like and why they're limited, and how governance works at personal, institutional, and political levels. This is not a credential — it's a lens. The next AI headline you read, the next chatbot you use, the next election that includes an AI policy platform: you will read all of it differently. That difference is real, and it's permanent.
A state legislator's office has reached out. They're drafting a bill about AI use in schools — specifically whether schools can use AI grading systems for student essays. They want a one-paragraph briefing from a young person who understands AI. Your lab partner will play a skeptical legislative aide who pushes back on your recommendations.
This lab is complete after 3 exchanges. You need to take a position and defend it with reasoning drawn from what you've learned — not just general feelings.