AIQ
AIQ
Everything a teacher needs to deliver this lesson — pick your grade's script below once you've read the background.
"AI Perception Challenge" is the capstone lesson of World 2. It doesn't teach a new idea — it asks students to take everything they've learned about how AI takes in the world (seeing, hearing, reading, and getting fooled) and use it to sort real examples correctly. By the end of the lesson, a student should be able to:
This lesson sits at the end of a five-lesson arc in World 2 ("How AI Sees & Hears"), and it only makes sense in that context. The three lessons before it each introduced one way that AI takes in information: Vision Explorer (2.1.1) covered how a computer "sees" a photo — not by looking at it the way a person does, but by reading it as a giant grid of numbers, one number per pixel for how bright or what color it is. Sound Detective (2.1.2) covered how AI "hears" — a microphone turns a voice into a sound wave, and speech-recognition software converts that wave into text. Text Decoder (2.1.3) covered how AI "reads" — natural language processing (NLP) turns words into numbers so a program can do math on them instead of understanding them the way a person does. Two more lessons filled in the picture: Data Collector (2.2.1) looked at the sensors that feed AI systems data in the first place, and Glitch Finder (2.2.2) looked at how each of those senses can be fooled — a sticker that confuses a self-driving car's camera, a made-up "fact" an AI states with total confidence, a joke an AI takes literally.
This lesson doesn't add a sixth idea. It's a sorting and synthesis exercise: given a real product — an Instagram filter, a spam folder, a voice assistant — can the student say which of those senses is doing the work? The payoff is the same for every age band: once a student can name "vision," "hearing," and "reading" as separate skills, they start noticing that the AI products they actually use rarely rely on just one. A voice assistant listens (speech recognition) and understands what you meant (language processing) and often has to decide what to say back (more language processing). That's the seed of the idea this lesson introduces by name for older students: multi-modal AI, meaning a single system built to take in more than one type of input together, rather than just one type at a time.
For older students, it's worth being precise about one thing: "multi-modal" doesn't mean an AI has five senses like a person. It means the software was built to process more than one type of data — image pixels, audio waveforms, and text tokens are fundamentally different kinds of data, stored and processed differently — and combine information from them to do a single task. A system that can only answer questions about a photo you show it (an image plus text) is doing something meaningfully harder than a system that only reads text, because it has to represent two very different kinds of information in a way it can compare and reason over together. For the oldest students, the lesson names two real systems that do this — CLIP, which learns to match images with their captions, and multi-modal chat models like GPT-4V or Gemini, which can take an image and a question about it in the same conversation.
One thing to actively avoid overstating, at every age: none of this means the AI "understands" sight, sound, or language the way a person does. Computer vision finds patterns in pixel numbers; speech recognition matches sound-wave patterns to likely words; NLP finds statistical patterns in text. They're powerful pattern-matching tools, not senses in the human sense, and the lesson's own "Glitch Finder" predecessor exists specifically to show how easily each of them breaks. Keep that thread alive here rather than letting the "wow, AI can do everything" framing stand unchallenged.
It also helps to know what this lesson is not trying to do, so you don't over-prepare. It isn't teaching students to build anything, and it isn't introducing new technical vocabulary for the younger bands at all — Explorer and Builder students work entirely with examples they've already met in the three earlier lessons of this world, just recombined into a sorting challenge. Only Hacker and Architect students meet a genuinely new label, "multi-modal," and even then it's naming something implicit in examples they've already seen (a voice assistant that both listens and understands) rather than introducing a new mechanism. If you're short on time, the lesson still holds together as a five-minute review game; the discussion questions and extension activities in each age-specific script are there to stretch it into a full period, not because the core content requires that much time.
The age-specific scripts linked below intentionally diverge more than most AIQ lessons do, because this is where the four age bands' different destinations become visible: a 6-year-old finishes this lesson able to say "that uses AI eyes," while a 17-year-old finishes it able to explain why CLIP's contrastive training objective lets it generalize to labels it's never seen. Both are the correct depth for the same underlying idea — that AI perception comes in distinct types, and combining them is where a lot of current capability (and a lot of current research effort) actually lives.
No special materials or advance prep are required. You need:
The lesson itself takes 5–15 minutes inside the app. Budget a full class period if you want to run the guided discussion and extension activity below alongside it.