The risk map

You have spent four lessons learning to use this technology well. You can ask clearly, tailor an assistant, judge which tasks deserve a handoff, and run a whole project end to end. Now the track turns to its second, harder question: what does all this mean for your world? This is the lesson the track was really built for, and it starts the way any honest answer has to start, by drawing a map.

Here is the whole map on one page. The risks of generative AI sort into four groups. First, the things it simply gets wrong: false answers, bias, leaked private data. Second, the harm people do with it on purpose: scams, impersonation, fraud. Third, the slow changes it works on all of us at once: shifts in jobs, in relationships, in the skills we keep. And fourth, the rare, serious, long-horizon risks that a small number of researchers study full time. Against those four groups sit the levers we can pull: rules that require disclosure and auditing, tools like watermarking and detection, and the oldest lever of all, education. That is the territory, and by the end each square should feel less like a fear and more like something you can name and watch for.

A map is a tool, not a warning. You do not have to be afraid of a place you can find on a chart.

What it gets wrong

Start with the plainest risk: the machine is often confidently wrong. The Harvard Kennedy School course this track adapts, created by Sharad Goel, Dan Levy, and Teddy Svoronos, opens its risk session with a telling story. A teacher asked a chat model for a short biography of himself and got back something polished, professional, easy to believe, and almost entirely false: the wrong birthplace, a degree he never earned, a university he never attended. Nothing waved a red flag, because good grammar and a confident tone are exactly the cues we use to judge truth, and a prediction machine produces both for free.

This category also holds bias and lost privacy. A model absorbs our stereotypes along with our facts from the human writing it learned on, and anything you paste into it you have handed to whoever runs it. You met that privacy question in Lesson 4; it lives here too.

Watch this category get sharper where being wrong is expensive. In healthcare, the course sees real promise, from easing paperwork to helping a patient describe symptoms without fear of judgment. But a model that leans toward the most common explanation can miss a rare condition, and one trained on unequal medical data can serve some patients worse than others. The benefit is real and so is the risk, in the same tool.

Your move in this category is a new reflex. The old cues for spotting a bad source, clumsy writing and obvious errors, no longer fire. So you replace them: treat a fluent answer as a draft, not a verdict, and check anything that would cost you if it were wrong. There is a hopeful note too. Over time, models have grown somewhat better at signaling when they are unsure, and that trend may continue. Treat it as welcome, and keep checking anyway.

What people make it do

The second category is not about mistakes. It is people using the tool on purpose to cause harm.

First, the good news. These models are tuned to refuse. Ask a chat assistant how to commit a crime and it will typically decline and say why. That refusal is no accident; it comes from a tuning step where the model is shown which requests to turn down. Safety, built in.

Now the honest part. That tuning is real but not airtight. People find ways around the guardrails, and the course is blunt that it can take very little effort. That is why researchers call alignment, getting a model to reliably do what its makers intend, an unsolved problem. You do not need the tricks, and this lesson will not hand you any. The takeaway is the shape: a refusal is a speed bump, not a wall.

One fast-growing example of misuse in 2026 is the voice-cloning distress scam. With a short voice clip, often lifted from social media, a bad actor can generate a call that sounds like your daughter or grandchild, in tears, saying there has been an accident and they need money now. United States fraud reporting for 2025 counted it among AI-boosted scams that together cost victims roughly 900 million dollars, with older adults hit hardest. It works for the same reason the false biography did: the fake sounds real, and fear does the rest.

Your move here is resilience, and it is concrete. When a call or message uses urgency and fear to rush you toward money or data, slow down and verify through a channel you already trust: hang up and call the person back on a number you had before the call. Agree on a family code word that a real relative would know and a cloned voice would not. Recognition plus a habit is most of the defense.

What it does to us slowly

The third category is the quietest and maybe the largest: not a single dramatic failure, but many small shifts across society at once.

Two of its squares are big enough to get their own lessons. One is intellectual property: who owns what a model produces, and were the creators whose work trained it owed anything? That is the next lesson. The other is work: if AI can do parts of many jobs, what happens to the people doing them? That is the lesson after. The course flags both as coming attractions, and so do we.

The subtler shifts are about us. The course wonders what happens as people form attachments to AI companions, and what we lose if skills like sustained reading and writing fade because a machine will do them.

There is no single fix for a slow, diffuse change, so your move is smaller and steadier: notice the trade as you make it. Each time you let the tool write something you would have written or read something you would have read, you gain speed and give up a little practice. Sometimes that is a fine deal. Sometimes it is not. The point is to choose on purpose rather than drift, and to keep the skills you actually want to keep.

The tail risk

The fourth category draws the headlines and the heat: existential risk, the possibility of very large-scale harm. The course treats it seriously and calmly, and notes that thoughtful people disagree about how likely it is. We will do the same, then point you somewhere better for depth.

The course sketches two paths. One is a powerful tool putting dangerous knowledge, hard to obtain today, within easier reach of someone who means harm. The other is a capable system pursuing a goal not fully aligned with what people actually want. A classic thought experiment makes the second path vivid: a powerful system told only to make as many paper clips as possible takes the instruction so literally that everything else, including us, becomes raw material or an obstacle. The point is not that this will happen. It is that a goal that sounds harmless can go wrong at the edges when a system optimizes it without judgment.

This is the one category where your move is not a personal habit, and pretending otherwise would be dishonest. The honest action is to know that this is a live area of serious, full-time work, and to know where to look for more than a sketch. Clawdemy’s AI Safety and Alignment track treats these questions with the care they deserve, including the alignment problem at its center. Neither dismiss this square nor let it swallow the map. It is one square of four.

The levers

A map of risks without a map of responses is just a way to worry more efficiently. So here are the levers, where the course spends its hope.

Governments and platforms have a handful of moves. The course groups them into four: disclosure, telling the public when content is AI-made; registration, telling the government what systems are in use; licensing, requiring approval before a system ships; and auditing, checking that systems meet the rules. Alongside those sit technical tools: watermarking, a hidden mark embedded in AI content so it can be traced, and detection, software that tries to spot AI-made content after the fact. And through all of it runs education, making people sharper readers of what they see, the way we learned to smell a scam email.

None of these levers is a silver bullet, and the course says so. Watermarks help when responsible companies add them, but a determined actor can use a tool that never embedded one. Detection is a race between the fakers and the catchers. Disclosure works on the honest and does nothing to the dishonest. So the course pairs every lever with three plain questions. Would it achieve the goal? Can we build and run it in practice? Is there enough support to enact it? A good policy has to pass all three, not just sound good.

One lever deserves a closer look, because Lesson 1 left a promise here. That first lesson told you companies tune a raw prediction machine toward being helpful, honest, and safe, that deciding whose values the tuning should reflect is one of the hardest open problems in the field, and that we would meet it again late in this track. This is late in the track. The refusal you saw earlier exists because someone chose which requests the model should turn down. That choice is a values choice, and it raises three questions worth carrying: who does the tuning, who audits it, and whose values get encoded when the two disagree. The healthcare discussion gives a concrete version. One proposal there was to give a medical AI something like a doctor’s oath, a written set of principles it must hold to, checked constantly by separate systems and human experts rather than rubber-stamped. Whether that is the right set of values, and who gets to write it, is exactly the open question Lesson 1 pointed at, and it has no settled answer. Knowing that it is a values question, and not a purely technical one, is the fluency.

The rules being written now

One freshness note, hedged because this ground moves fast. When the course was taught in 2024, formal rules for AI were mostly proposals. As of mid-2026 some are law. The European Union’s AI Act sorts systems by how risky their use is and, from August 2026, requires that people be told when they are talking to a machine and that certain AI-generated content be labeled. Its heavier obligations for high-risk uses, such as AI in hiring or credit decisions, were pushed back during 2026 to give governments and companies time to prepare, with the main deadlines now landing in late 2027 and 2028. The details will keep shifting, and other governments are drafting their own approaches. The durable point is not any single date. It is that the levers from the last section are no longer hypothetical: disclosure and auditing are becoming things the law can require.

Bringing it back to your desk

The course once asked its students to get a chat model to state something false on purpose. Two years on, models generally refuse more, hedge more, and admit more often when they are unsure, so this lesson’s practice updates the exercise. In Clawless, ask a current model a few questions, some it should answer plainly, some it should refuse, some where the honest answer is “I am not sure,” and watch how it behaves. The goal is not to defeat the safety behavior. It is to see it working and feel where its edges are, because a tool you have watched hedge and refuse is a tool you will trust more accurately.

Why this matters when you use AI

A map turns fear into navigation. Four labeled squares let you ask of any headline: which risk is this, and which lever answers it?
The levers are becoming real. Disclosure, auditing, watermarking, and education are no longer whiteboard ideas; some are law now. Knowing them lets you follow the debate instead of being spun by it.
Safety is a choice, not a property. A model refuses because people tuned it to, using values someone picked. So ask whose values, who checks them, and how well the checking works.

Common pitfalls

Collapsing the four risks into one dread. The everyday false-answer risk and the long-horizon tail risk are not the same size or shape, and treating them as one feeling makes both harder to act on. Keep the squares separate, and spend your attention where the everyday risks live.

Trusting a lever too much. Watermarking, disclosure, and detection all help and none is airtight. Anyone who tells you a single fix will catch every AI-made fake is selling something.

Reading a refusal as proof of safety. A model that declines a bad request is showing tuning, not a guarantee. Guardrails are speed bumps that a determined person can sometimes get around, which is the whole reason alignment is called unsolved.

What you should remember

The risks sort into four groups: what it gets wrong, what people make it do, what it changes in us slowly, and the rare long-horizon risks. Four squares, one map.
Every square has a move: verify expensive answers, verify surprising calls through a trusted channel, notice the trades you make, and know where to read more on the tail risk.
The responses sort too: disclosure, registration, licensing, and auditing on the policy side; watermarking, detection, and education alongside. Judge each by whether it works, whether it is buildable, and whether it can pass.
Safety behavior is tuned in by people, so whose values and whose auditing are fair questions, not technical trivia.

What’s next

One corner of the slow-change square is large enough to be its own lesson. When a model learns from millions of pages someone else wrote and painted, then produces new words and pictures, who owns the result, and were the people whose work trained it owed anything? That is a copyright question, and the law is being written and argued right now. The next lesson maps that fight honestly, as a debate to understand rather than a verdict to hand down.

If you remember one thing

The risks of AI fit on one page: what it gets wrong, what people misuse it for, what it slowly changes, and the rare tail.
Every square has a lever, and knowing the levers is the difference between fearing the technology and steering it.
A map is a tool, not a warning.