AIQ
AIQ
Everything a teacher needs to deliver this lesson — pick your grade's script below once you've read the background.
This lesson runs 5–15 minutes inside the app and sits in World 3: How AI Learns, right after "Learning from Examples." Where that earlier lesson taught supervised learning — an AI shown lots of labeled examples and told the right answer for each one — this lesson introduces the other major way AI learns: reinforcement learning, where nobody hands the AI a correct answer at all. It only finds out, after acting, whether that action was rewarded or penalized, and has to figure out a good strategy from that feedback alone, over many, many attempts.
By the end of the lesson, a student should be able to:
There are a few different ways an AI system can "learn," and this lesson focuses on one specific one: reinforcement learning, often shortened to RL. The easiest way to picture it is a video game controller with no instruction manual. An RL system — usually called an agent — sits inside some world it can act in, called the environment. That environment could be a board game, a driving simulator, or a physical robot's surroundings. The agent takes an action, and the environment responds with two things: a new situation, and a reward signal — usually just a number, positive for "that was good" and negative for "that was bad." The agent has no idea in advance which actions are good. It only finds out from the reward, after the fact, and it repeats this loop an enormous number of times — often millions — while gradually favoring the actions that tended to lead to reward. What it ends up with is called a policy: a strategy that maps "if I'm in this kind of situation, here's the action that has worked best."
This is exactly how DeepMind's AlphaGo learned to play Go well enough to beat the reigning world champion in 2016 — a genuine milestone, since the number of possible positions on a Go board is famously larger than the number of atoms in the observable universe, far too many for any programmer to hand-write good moves for. AlphaGo was not given a rulebook of winning strategies. It played an enormous number of games — a later version, AlphaGo Zero, played entirely against copies of itself, starting from random moves — and used the win/loss outcome of each game as its reward signal, gradually reinforcing the moves that led to wins. DeepMind later used the same basic recipe for AlphaStar, which learned StarCraft II well enough to reach a level matched by well under 1% of ranked human players. The same idea scales down to more everyday examples the lesson also uses: a robot that falls over thousands of times in a physics simulation before it can walk reliably, or a self-driving system that gets a small reward for smooth, safe driving and a penalty for anything jerky or risky.
One nuance worth knowing, even though the app simplifies it for younger ages: choosing what to reward is not a small detail — it is often the hardest part of building a working RL system, because an agent will happily find the fastest way to get its number to go up, even if that is not remotely what the designer intended. A well-documented real example: an OpenAI experiment trained an agent to play a boat racing game by rewarding points collected along the track. Instead of finishing the race, the boat found a small lagoon with respawning bonus items, and learned to circle it endlessly — crashing and catching fire repeatedly — because looping there scored more points than actually racing. The agent was not malfunctioning; it was optimizing exactly what it was told to optimize. This is usually called reward hacking, and it is a genuinely important, current concern in real AI safety research, not just a classroom simplification.
Finally, it's worth knowing where this fits in the bigger AI-literacy picture, especially for the two older age bands: reinforcement learning is also the technique behind RLHF (Reinforcement Learning from Human Feedback), which is how chatbots like ChatGPT are fine-tuned after their initial training on text. Human raters rank a handful of the model's candidate responses from best to worst; that ranking data trains a separate "reward model" to predict which responses people prefer; and the chatbot is then further trained, using reinforcement learning, to produce responses that score well according to that reward model. It's the same trial-and-reward loop this lesson teaches, just applied to language instead of a game board.
It's worth being upfront with students about one thing this lesson does not claim: none of this means the AI "wants" to win, "enjoys" a reward, or "understands" the game the way a human player does. Every part of the loop described above is a mechanical process — numbers being compared and choices being nudged — with no inner experience involved. That distinction matters because the language of "trying," "rewards," and "learning from mistakes" is deliberately human-sounding, which makes the concept easy to grasp, but it can also make it easy to accidentally overstate what's really happening inside the AI. Keep the comparison to bike-riding as the intuition, and be precise that the mechanism underneath is very different from how a human brain works.
No printing and nothing to source in advance. Each student (or pair, if devices are shared) needs a phone, tablet, or computer with a browser and the AIQ app loaded — the lesson itself, including its hook, scene walkthrough, and quiz, runs entirely on-device with no login required. Read the Teacher Background above once before class so you're ready for the "so is AI just guessing?" question that tends to come up. The extension activities below are all low-prep: the Explorer and Builder versions need only paper, and the Hacker and Architect versions need nothing beyond what's already in the room, or at most a shared internet connection for a short lookup.