The risk map: In brief
Lesson 6 turns the track from its first question to its harder second one. The first four lessons taught you to use generative AI well, to ask clearly, tailor an assistant, decide which tasks deserve a handoff, and run a whole project end to end. 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. It adapts the risk session that opens the implications unit of the Harvard Kennedy School course this track is built on, keeping that session’s calm, sorted structure: four kinds of risk, then the levers for managing them.
The map is the whole point. The risks of generative AI sort into four groups, and once they are sorted, each becomes something you can name and watch for rather than a single cloud of dread. First, the things it simply gets wrong: false or fabricated 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, where thoughtful people disagree about how likely it is. A map is a tool, not a warning.
The capability: after this lesson, you can take any AI risk you meet in a headline and place it on the four-category map, then name the lever or personal move that answers it; you can recognize a voice-cloning distress scam and carry the resilience habit that defuses it; you can judge a proposed policy by whether it would work, whether it is buildable, and whether it can pass; and you can see why a refusal is a tuned choice made by people, not a property the machine simply has, which turns whose values and whose auditing into fair questions rather than technical trivia.
What the lesson covers. First the four squares, each closing on a reader action. What it gets wrong: a fluent answer is a draft, not a verdict, so verify anything that would cost you if it were wrong, and note the hopeful trend that models have grown somewhat better at signaling when they are unsure, and that this trend may continue. What people make it do: safety tuning is real but not airtight, which is why alignment is called an unsolved problem, and one fast-growing example is the voice-cloning distress scam, counted in United States fraud reporting for 2025 among AI-boosted scams that together cost victims roughly 900 million dollars. What it changes slowly: attachments to AI companions and the fade of skills we stop practicing, so notice the trade as you make it. And the tail risk, treated seriously and calmly, with depth cross-linked to the AI Safety and Alignment track rather than re-taught here. Then the levers: disclosure, registration, licensing, and auditing on the policy side; watermarking, detection, and education alongside; each judged by three plain questions. A freshness note lands here too: as of mid-2026 some of these levers are law, with the European Union’s AI Act requiring disclosure of machine interaction and labeling of certain AI-generated content, and its heavier obligations for high-risk uses now landing in late 2027 and 2028.
Why this order. The categories run from the risk you meet daily to the one a few people study full time, so the reader spends most attention where the everyday risks live and meets the tail last, as one square of four. The levers section follows because a map of risks without a map of responses is just a way to worry more efficiently, and it pays back a promise from lesson 1: the tuning that makes a model helpful, honest, and safe encodes someone’s values, and this is where the track said it would return to who chooses them. The practice runs two exercises in Clawless: a sorting drill that places plain scenarios on the map, and a transformed observation exercise where you watch a current model answer, refuse, and admit uncertainty, to see the safety behavior working. Lesson 7 opens one corner of the slow-change square into its own lesson, the copyright question of who owns what a model makes and learns from, offered as a debate to understand rather than a verdict to hand down.