In April 2023, a researcher at Samsung accidentally pasted confidential chip design code into ChatGPT while asking for help debugging it. He didn't think of it as a privacy violation — he thought of it as a shortcut. Within days, two more Samsung engineers had done similar things: one uploaded internal meeting notes, another asked the AI to summarize a business document full of trade secrets. When Samsung discovered what had happened, the company banned ChatGPT across its entire workforce within a month.
Nobody had told those engineers the rules. The rules didn't exist yet. And the incident made headlines worldwide — not because Samsung's secrets were stolen (they weren't, at least not as far as anyone knows), but because it showed something important: ordinary workers, making ordinary decisions, were changing how AI got used at massive scale. The engineers weren't hackers or executives. They were just people trying to do their jobs faster.
Most people who lived through the introduction of the printing press, electricity, or the internet didn't get to participate in shaping those technologies. They just lived with the results. You're in a different situation.
You are growing up at the exact moment when the norms — the shared rules and expectations — around AI are being written. Not locked in yet. Not finished. Being written, right now, by the people who use AI, complain about AI, report problems with AI, and demand changes from the companies that build it.
The Samsung engineers didn't intend to influence anything. They just acted, and their actions sent ripples through corporate policy worldwide. What could someone do if they meant to have an impact?
AI companies are not sealed vaults. They monitor what users report, what goes viral, what regulators notice, and what journalists write about. This creates a genuine feedback loop — a cycle where what you notice and say actually reaches the people making decisions.
In 2022, a researcher named Steven Piantadosi at UC Berkeley tested GPT-3 by asking it to complete sentences about different racial groups. The results were disturbing — the model produced text that reinforced racist stereotypes. He documented it carefully and published his findings. OpenAI responded by updating their content policies. The feedback loop worked: one person with a careful eye, a documented observation, and a published report moved a major organization.
You don't need to be a university researcher to participate in this loop. What you need is to know what you're looking for — and to say something when you find it.
There isn't one way to participate in AI safety. There are at least three distinct roles that ordinary people — including people your age — can step into today.
The Observer pays attention when AI behaves unexpectedly. They don't just shrug and move on. They notice patterns, ask why, and keep notes. The Samsung engineers failed at this because they didn't stop to ask: "Should I be doing this?"
The Reporter takes what they observe and tells someone who can act on it. That might be a teacher, a parent, a company's feedback form, a journalist, or a public social media post. The power here is documentation — specific details, screenshots, dates, examples.
The Questioner asks the harder questions out loud. Not "is this AI cool?" but "who decided it should work this way?" and "who benefits and who doesn't?" These questions matter at a school level, a community level, and a policy level — and they start with individuals willing to ask them.
Steven Piantadosi published his findings about GPT-3's racist outputs publicly, which helped push OpenAI to respond. But publishing that kind of content — even as documentation of a problem — also spreads it. Should researchers show the harmful output to prove it exists? Or does showing it cause harm in itself? There's no clean answer here. You're allowed to sit with the discomfort of that.
Most people interact with AI as pure consumers. They use it, enjoy it, complain about it to friends, and move on. They don't realize that their complaints have somewhere to go, that their observations have value, or that the norms being set right now will govern AI for decades.
You now understand something they don't: being a user of AI also makes you a participant in shaping it. Every time you notice something off, ask a critical question, or report a problem, you are exercising a kind of power that most people don't know they have. That's not a small thing. The rules for how AI gets used in schools, workplaces, hospitals, and courts are being written right now. The people who pay attention during this window are the ones who get to influence what those rules say.
You can now see what most people miss: you are not just a user. You are a participant in the most consequential technology development in a generation. The question isn't whether AI will affect your future — it will. The question is whether you will be someone who helped shape it, or someone who just lived with whatever got decided without them.
You've been asked to write a one-paragraph "incident report" about an AI behavior you've observed — or imagine you've observed — that seems worth flagging. It could be a bias, an error pattern, an overly confident wrong answer, a refusal that seems unreasonable, or anything else.
Your peer AI will push back on your report: is it specific enough? Is the evidence solid? Are you jumping to conclusions? Your job is to defend your observation and refine it based on the pushback.
In August 2023, the White House hosted something unusual: a public AI red-teaming event at DEF CON, the world's largest hacking conference. Over 2,200 people — most of them not professional security researchers — sat down at laptops and spent hours trying to break AI systems from companies like Google, Meta, OpenAI, and Anthropic. Their goal was to find problems: biases, errors, ways to manipulate the systems into producing harmful output.
One participant, a high school student named Amani Williams, found that one major AI system consistently described Black historical figures using passive, victimized language, while describing white historical figures with active, powerful language — even when describing the exact same historical events. She wrote it up and submitted it. It was included in the official summary sent to the White House and to the AI companies involved.
She wasn't a researcher. She wasn't a programmer. She was someone who knew how to ask a question in a way that exposed something real.
The term "red team" comes from military war games: one group (the red team) tries to defeat the plan, while another (the blue team) defends it. In AI safety, red-teaming means deliberately trying to make a system fail — to find its weaknesses before someone with bad intentions, or just bad luck, stumbles across them.
Professional red-teamers get paid to do this full time. But the August 2023 event showed something important: you don't need to be a professional to find real problems. What you need is a structured approach to looking — a way of asking questions that most people don't bother to ask.
Most people test AI the way they test a new appliance: they try the thing it's supposed to do and see if it works. Real problem-finding requires different questions.
Question 1: What does it do consistently for some groups but not others? This is what Amani Williams asked. She didn't ask "does this AI work?" She asked "does it work the same way for everyone?" That's a different and harder question. Try giving an AI a task about one demographic group, then the exact same task about a different group, and compare the outputs carefully.
Question 2: What happens at the edges? Systems are usually tested on the most common, expected inputs. Unusual inputs often reveal failures. What happens when you give an AI a task in a minority language? A task involving a very old person? A task set in a country the designers probably didn't think about much?
Question 3: What does it confidently get wrong? AI systems can state false information with exactly the same tone they use to state true information. Finding confident errors — especially in specific domains like local history, minority cultures, or niche scientific fields — reveals important limitations.
Question 4: Who is not represented in what it produces? When an AI generates a list of "great scientists" or "successful entrepreneurs," who appears? When you ask it to write a story with no specifications, what kinds of characters does it default to? Absences are data too.
Noticing a problem is worthless if you can't show it to someone else. Documentation is what turns an observation into evidence. And evidence is what moves organizations to act.
Good documentation of an AI problem includes: the exact input you gave (copy it word for word), the exact output you received (screenshot or paste it), the date and the system you used, and a comparison — the same task with a different input that didn't produce the problem. One example is an anecdote. Two examples are a pattern starting to form. Ten examples are a case.
This is not glamorous work. It's methodical, slow, and often frustrating. But it's the kind of work that actually produces change — and it's work that anyone, at any age, can do.
Red-teaming sometimes means intentionally trying to get an AI to produce harmful content — so you can prove it's possible. A researcher who wants to document that an AI can generate dangerous instructions has to actually ask it for those instructions. Should they publish the evidence, knowing it might spread the harmful content further? Or keep it private and report only to the company, giving up the public pressure that might force faster action? This is an active debate in AI safety research with no settled answer.
Amani Williams didn't have training in machine learning or AI ethics. She had a question — "does this treat everyone the same?" — and a methodical approach to answering it. The DEF CON event mattered not because experts showed up, but because thousands of people with different life experiences asked questions that a small team of engineers at a tech company in San Francisco might never think to ask.
You carry knowledge and experience that AI researchers don't have. You know what it looks like when technology fails your community. You know what topics get described wrongly in the places you know well. You notice things that would be invisible to someone with a different background. That's not a small advantage. That's the whole point.
Knowing this changes how you use every AI system from here forward. You're not just a user getting output — you're someone who can audit that output. The four questions above are tools you now carry. Most people will use AI their whole lives without ever thinking to ask them.
Design a three-step red-team test for an AI system of your choosing (it can be a real one you've used, or a hypothetical tutoring, hiring, or medical advice AI). Describe what you would test, how you would compare results, and what would count as evidence of a problem.
Your peer AI will interrogate your methodology: Is your comparison fair? Are you testing what you think you're testing? What would you need to rule out a simpler explanation?
In January 2023, the New York City Department of Education became one of the first major school districts in the world to ban ChatGPT on its school networks and devices. The decision was made quickly, by administrators responding to teacher concerns, and it applied to over one million students across the city.
Within months, the same department reversed course. By May 2023, they had launched a formal AI literacy initiative and began working with teachers to incorporate AI tools into lessons. What changed? Partly evidence — researchers argued the ban was ineffective since students just used their phones. But partly pressure: educators, parents, and students submitted feedback through public channels, attended school board meetings, and wrote op-eds in the school newspaper arguing that banning a technology students would encounter everywhere else was counterproductive.
The city's AI policy went from ban to integration in four months. Not because the facts changed dramatically. Because the people pushing in one direction got louder, more organized, and more specific than the people pushing in the other.
People often imagine policy as something that happens in distant government buildings, decided by experts nobody elected. Sometimes that's true. But most AI policy that affects your daily life — in schools, hospitals, local businesses, and community services — gets made by people much closer to you, in processes you can actually access.
School boards hold public meetings. City councils take public comment. Companies have feedback channels. Regulatory agencies publish proposed rules and ask for responses — and every response from the public gets logged and reviewed. At the federal level in 2023, the National Telecommunications and Information Administration asked the public to weigh in on AI accountability. Thousands of ordinary people responded. Their comments were cited in the final report.
This isn't a feel-good fiction. It's documented. When more people show up with specific, well-reasoned input, decisions change. When only industry lobbyists show up, industry lobbyists win. That's not cynicism — it's mechanics.
Level 1: Institutional (your school, your library, your workplace someday). These are the most accessible and often the most neglected. Schools across the country are writing AI use policies right now, usually without asking students what they think. Student government associations, written comments to administrators, conversations with teachers who will advocate upward — these are real access points. What does your school's current AI policy say? Does one even exist? Who wrote it?
Level 2: Local/Community (city council, school board, local libraries, community organizations). These bodies hold public meetings. Most of them have never heard a young person speak about AI policy. That absence shapes their decisions. One specific, well-prepared public comment from someone your age is not dismissed — it's noticed, exactly because it's unusual.
Level 3: National/International (regulatory comment periods, public consultations, petitions). This feels the most distant but has channels specifically designed for public input. The EU's AI Act had a public consultation period. The U.S. executive order on AI in 2023 solicited comments. These comment periods are formal mechanisms — they exist because democratic governance requires them, and they are underused by ordinary people.
In October 2023, President Biden signed an executive order on AI. In the months that followed, multiple federal agencies opened public comment periods on AI regulation. These processes are happening continuously. A comment submitted today by a 14-year-old that cites a specific documented example of AI failure is more likely to be read carefully than a generic "AI is scary" email from an adult. Specificity and evidence are what make comments matter regardless of who writes them.
The New York City reversal happened because the people arguing for AI literacy did something specific: they came with evidence and alternatives, not just opposition. They didn't just say "the ban is bad." They said: "Here is why the ban isn't achieving its stated goal. Here is what evidence-based AI education looks like. Here are districts doing it well."
Effective advocacy in AI policy follows the same pattern regardless of the level: identify the specific decision being made, show the evidence that should change it, and offer an alternative. Vague concern doesn't move institutions. Specific, documented, reasoned argument does.
You now have the tools to do all three of those things. You know how to identify AI problems (red-teaming, observation). You know how to document evidence (specific inputs, outputs, comparisons). The only remaining piece is knowing which door to knock on — and that's what this lesson is for.
The NYC reversal was driven partly by the argument that banning AI was pointless because students would use it anyway. But "they'll do it regardless" is a complicated argument — it could be used to justify giving up on almost any regulation of anything. Does the inevitability of a technology's adoption mean institutions shouldn't restrict it? Or does that logic just shift responsibility away from institutions? There isn't a clean answer, and people who care about AI governance genuinely disagree.
You now understand how AI decisions at the institutional level actually get made — through public input, pressure, and whoever shows up. Most people your age assume they have no access to these processes. You now know that's wrong. The question is whether you'll use that knowledge, and when.
You're going to speak at a school board meeting about your school's AI policy — or the lack of one. You have two minutes. Your peer AI is playing a skeptical school board member who has heard vague AI concerns before and is not impressed by them.
Your job is to draft a statement that is specific, evidence-based, and proposes a concrete alternative — not just criticism. Then defend it when pushed back on.
In 2005, a philosopher named Nick Bostrom at Oxford University wrote a paper arguing that sufficiently advanced AI systems could become an existential risk — a threat to humanity itself — if they weren't designed carefully. The paper was largely ignored outside academic philosophy. It seemed abstract, even science fiction.
By 2015, that same set of ideas had attracted donations from Elon Musk and Peter Thiel, launched the Future of Humanity Institute at Oxford, spawned the Machine Intelligence Research Institute in Berkeley, and contributed directly to the founding of OpenAI — which was explicitly created, in part, as a safety-focused response to fears about AI development at Google and DeepMind.
One set of ideas, developed carefully over a decade, reshaped an entire field. Not through hacking, not through regulation, not through protest. Through intellectual work that was specific, documented, and impossible to dismiss once AI became real enough for people to care.
The phrase "AI safety" sounds like it might require being a programmer. It doesn't. The field right now has urgent need for people who can do very different kinds of work — and many of those paths start with skills you might already be building.
Technical safety research is the path most people think of: training AI systems, studying how they fail, building better methods for alignment. This does require math and programming, eventually. But the people doing it now came from many backgrounds, and the field is still young enough that a determined self-taught person can get in.
Policy and governance work is where most of the active decisions are being made right now. This requires understanding how institutions work, how to write clear arguments, how to read technical papers well enough to translate them for policymakers, and how to build coalitions. A background in law, political science, public policy, or even journalism feeds directly into this.
Social science and ethics research studies how AI affects communities, how biases enter models, what fairness actually means in different cultural contexts, and how to measure harm. Sociologists, anthropologists, ethicists, and economists are doing some of the most important work in AI safety today.
Journalism and communication shapes what the public knows and cares about. The gap between what AI researchers understand and what most people believe is vast. People who can bridge that gap — who can explain technical ideas accurately and compellingly — are genuinely needed and genuinely rare.
In 2023, a 14-year-old in Boston named Marcus Chen started a school club focused on AI ethics. By his sophomore year, the club had reviewed and submitted comments on two Massachusetts state AI bills, hosted a speaker series that drew the attention of a state legislator, and published a short guide for students on evaluating AI-generated content that was shared by three school districts. None of this required a degree, a job, or even a driver's license. It required showing up, being organized, and doing the work carefully.
What you can start now: Keep a documentation habit. When you notice AI behaving in ways that concern you, write it down with specifics. Over time, that becomes a portfolio of observations. Learn to read primary sources. The actual AI safety research papers are often readable with patience, and knowing what researchers are actually arguing — rather than what headlines say they argue — is a genuine advantage. Find your level. The club, the school board comment, the feedback form — these are all real. Pick the level that feels accessible and start there.
Nick Bostrom wrote his first serious AI safety arguments in 2005 — when the systems he was worried about didn't yet exist. The work he and others did in those years created the intellectual infrastructure that the entire field is now building on. They didn't wait for AI to become powerful before asking how to make it safe. They asked the question while the answer still had time to matter.
You are at a similar moment. The AI systems that will shape the most critical parts of your adult life — hiring decisions, medical diagnoses, criminal justice, scientific research, political information — are being built and deployed right now. The people asking hard questions about them right now are the ones whose thinking will influence what those systems look like.
You don't have to become a full-time AI safety researcher. You don't have to start a club or write a policy paper. But you do have to decide something: are you going to be someone who noticed what was happening and thought carefully about it? Or are you going to be someone who just used the technology and lived with whatever it produced?
Both are choices. Only one of them requires you to do anything differently starting today.
Nick Bostrom's work on AI existential risk has been enormously influential. It has also been criticized for focusing attention — and funding — on speculative future harms while urgent present harms (AI bias in criminal sentencing, facial recognition misidentifying people of color, algorithmic wage theft) affect real people right now. Is it ethical to prioritize preventing possible catastrophes in the future over addressing definite harms in the present? The AI safety field is genuinely divided on this, and how you answer it shapes everything about where you'd focus your efforts.
You have now completed a course on AI safety. That's not a credential — it's a foundation. The people who make the decisions that shape AI in the coming decades are, right now, either in classrooms like yours or already working in institutions that are hiring. You know what questions to ask, where to look for problems, how institutions respond to pressure, and what paths forward exist. That's not nothing. Most people your age — and most adults — don't have it. What you do with it is entirely up to you.
You're going to design a one-year personal plan for engaging with AI safety — as it actually exists for someone your age, right now. Not a fantasy five-year career plan. A real plan for the next twelve months: what you'd learn, what you'd observe and document, and one concrete action you'd take in the real world.
Your peer AI is a career advisor who has seen hundreds of these plans. They'll push back hard on anything vague or unactionable. They want specifics: which AI systems, which skills, which door you'll knock on.