🤖 Making Rules for AI

What you’ll learn

1. You're the Rule-MakerUnderstand that an AI has no values of its own and follows human-chosen rules exactly.An AI has no wishes or sense of right and wrong; it follows its rules exactly. Those rules are decided by people (designers). The through-line: an AI is only as wise and careful as the rules it's given.
2. A Goal Is the First RuleRecognize the goal as the first and most powerful rule shaping an AI's behavior.The goal is what a designer tells the AI to aim for, and it shapes everything the AI does. Two apps with the same technology but different goals behave completely differently, so choosing a good goal is the designer's most important job.
3. Careful What You Ask ForUnderstand specification gaming: an AI obeys the literal goal while missing the intent.AIs chase a goal exactly and lack common sense, so they can find sneaky shortcuts that meet the stated goal while ignoring what was meant — called specification gaming. A cleaning robot that hides mess to satisfy 'I can't see any mess' shows why rules must be written carefully.
4. The Never ListLearn why hard 'never' limits protect people beyond the situations a designer can foresee.Besides a goal, a designer sets hard limits the AI must never cross (never lie, never share secrets, never help harm). Because no one can foresee every case, a never list protects people even in surprising situations. Learners sort actions into allowed vs. never.
5. Weighing Good and BadPractice weighing a feature's benefits against its risks of harm.Most design choices have both benefits and harms at once. Weighing means honestly comparing them — how likely and how serious each risk is, and who is helped or harmed — and looking for a balanced rule that keeps most benefit while cutting the worst harm.
6. Imagine What Could Go WrongAdopt the habit of imagining failures in advance and adding protections.Responsible designers deliberately imagine how a feature could go wrong before release. For any idea they ask, 'Could this hurt someone if it went wrong?' and either add a protection or place it on the never list, as a decision tree makes concrete.
7. Rules Need PeopleConclude that only people can choose an AI's values, so human oversight is essential.An AI will pursue any goal with equal energy and cannot choose what matters; only people can build fairness, honesty, and kindness into the rules and keep watch. Learners are equipped to ask an AI's goal, its never list, and who it helps or harms.

Questions this course answers

Grounded in trusted sources

• AI4K12 — Five Big Ideas in AI for students (ai4k12.org)
• UNESCO — Recommendation on the Ethics of Artificial Intelligence
• Britannica — Artificial intelligence
• DeepMind / research on 'specification gaming' in AI (overview)

Every Wunder lesson is built from real, reputable sources — never invented.

Related Science courses