Teaching "How Computers See" to Hacker mode (ages 11–14)

Hook & Warm-Up

This age band gets its own hook, aimed at the mechanism rather than the "wow" factor. Read it close to verbatim, then open it up for a guess before revealing anything more:

"How does AI actually see? It turns images into numbers, then uses special networks to find patterns. Let's explore the journey from raw pixels to recognizing objects."

Follow with: "I want a guess before I explain anything. If I hand a computer a photo as nothing but a giant grid of numbers — no labels, no context — what's the FIRST, simplest thing you think it could possibly notice about that grid?" Push toward: places where numbers change sharply next to each other (that's an edge). Whatever answer they give, tell them they'll find out in a minute whether they're close — this is exactly how the real system works, building up from simple to complex.

Follow up with a second framing question that sets up the whole lesson: "A photo of a cat and a photo of a dog might both have four legs, fur, two eyes, a tail — lots of overlap. If you were writing rules by hand (like 'if it has pointy ears, it's a cat'), how many rules do you think you'd need before you covered every real photo out there?" Let students realize the number would be enormous and full of exceptions. This is exactly why modern computer vision abandoned hand-written rules in favor of learning patterns from labeled examples — a shift that's core to the concepts this age band's quiz will test.

Main Activity

Frame the three scenes as a pipeline — pixels in, understanding out — since that framing is what this age band's hook and summary are pointing at.

Scene 1 — Pixels (the raw input)

Cover the base facts: a pixel is one tiny colored dot, built from Red, Green, and Blue values mixed together; a digital photo is a grid of numbers where each number sets one pixel's color; zoom into any image far enough and the squares appear, because that's genuinely all a computer image is, at every zoom level. Reinforce that a computer never has anything richer than this grid to work with — no notion of "edge," "face," or "object" exists until software specifically looks for it.

Make the scale concrete: a modest 1000×1000 pixel image, with three color values per pixel, is three million individual numbers before any analysis begins. A neural network doesn't examine that grid the way a person scans a photo — it performs a sequence of mathematical operations across the entire grid at once, and those operations are what the "layers" in the next scene actually are. It's worth being upfront that this is genuinely a lot of computation, which is part of why specialized computer chips (GPUs, and increasingly dedicated AI chips) exist — ordinary processors can do this work, just much more slowly.

Scene 2 — Face Detection (a worked pattern-matching example)

Explain that AI scans for the general pattern of two eyes, a nose, and a mouth arranged the usual way, then measures distances between features — eye spacing, nose width, jaw shape — to build a numeric "face fingerprint" stored and compared each time you unlock your phone. Use this as your entry point into the age band's summary content:

"AI vision works in layers. First layers spot edges. Middle layers find shapes. Deep layers recognize whole objects. Face recognition maps the distances between your eyes, nose, and mouth to create a unique face template."

Unpack "layers" concretely: a system built this way is called a convolutional neural network. Early layers are only sensitive to simple things like edges and color changes; each following layer combines the previous layer's findings into something more complex — edges become corners and curves, then those become shapes like "circle" or "triangle," and eventually shapes combine into recognizable objects like "eye" or "wheel." No single layer "knows" what a face is — the recognition emerges from the combination.

It helps to draw this as a literal funnel on the board: a wide box labeled "raw pixels" narrowing through boxes labeled "edges," then "shapes/textures," then "parts (eye, nose, wheel, leaf-edge)," then finally "whole object" at the narrow end. Emphasize that the network isn't told what an "edge" or a "shape" is in advance — it discovers, during training on labeled examples, which patterns of numbers are useful for telling categories apart. Nobody hand-writes "an edge is where pixel values change by more than X" into the final trained system; the network arrives at something functionally similar on its own, purely from being shown enough labeled examples and being nudged, example by example, toward better guesses.

Scene 3 — Object Recognition (stakes and scale)

Cover: self-driving cars recognizing cars, pedestrians, and signs from camera images; AI spotting disease patterns in X-rays and MRIs; Amazon Go's cashier-less checkout cameras; and plant-ID apps like PlantNet. Then bridge into the resolution/data-quality quiz question by asking: "Between a really sharp camera and a really large, varied set of training photos, which do you think matters more for accuracy? Why?" — steer toward training data quality and diversity being the bigger factor, since a model can only recognize variation it has actually been trained on.

This is a good moment to introduce why self-driving cars specifically don't rely on cameras alone. Ask: "A camera can be blinded by glare, fog, or darkness. What might engineers add to make the system more reliable?" Lead toward lidar (which measures distance using laser pulses) and radar (which detects motion and distance using radio waves) — both feed into the same overall system alongside cameras, a practice called sensor fusion, precisely because no single sensor is reliable in every condition a car might drive through.

For the medical-imaging example, it's worth being precise about what "AI can spot diseases" actually means in practice: these systems are typically decision-support tools that flag areas of a scan worth a closer look, used alongside a radiologist's own reading, not full replacements making unsupervised diagnoses. This matters for accuracy in your teaching — overstating AI's autonomy here is a common and easy mistake to make, and it's worth correcting if a student assumes otherwise.

Have students complete the lesson individually on their devices, including the practice round classifying tasks as AI vision or not.

Discussion

• "Why would a convolutional neural network need to learn edges before it can learn to recognize a whole face?"
• "A self-driving car needs to recognize objects in real time, while a plant-ID app can take a couple of seconds. What tradeoffs might engineers make differently for each?"
• "If face recognition is based on measurements, not stored photos, what happens (in theory) to identical twins with very similar face geometry?"
• "Medical AI reads X-rays looking for patterns doctors might miss. Should a doctor ever trust an AI's answer without checking it themselves? Why or why not?"
• "What kinds of training photos would a self-driving car's vision system need that a house-cat identification app would not?"
• "Sensor fusion means combining cameras, lidar, and radar so no single sensor's weakness sinks the whole system. Can you think of a non-vehicle example where combining multiple 'senses' or data sources would make an AI system more reliable?"

For this age band, push back gently on confident one-word answers and ask "why?" a second time — the goal of this discussion isn't just recalling facts from the lesson, it's practicing the habit of asking how an AI claim would actually be verified, which is a skill this age group is ready to start building.

Quiz Walkthrough

This age band gets its own quiz, separate from the base lesson's — it uses more precise technical vocabulary (convolutional neural network, biometric template, sensor fusion) and is designed to check whether the layered edges-to-objects idea actually landed, not just whether students remember the examples.

Wrap-Up & Extension

Close with the age band's own summary: "AI vision works in layers. First layers spot edges. Middle layers find shapes. Deep layers recognize whole objects." Add: "Every computer vision system we talked about today — face unlock, self-driving cars, medical imaging, plant ID — is a version of this same layered idea, just trained on different data for different jobs."

Extension activity: Have students research and briefly present (2–3 minutes each, or as short written summaries) one real-world case where computer vision got something wrong — a well-documented example of a self-driving car misreading a sign, a face-recognition misidentification, or a medical-imaging false positive/negative. Ask them to identify, from what they now know about the pixels-to-patterns pipeline, a plausible reason the system might have failed (bad lighting, a rare/unusual object not well represented in training data, ambiguous edges). This connects today's mechanism to why "AI didn't understand what it saw" has real consequences.

A second extension, better suited to a class with some coding exposure: point students to any free, browser-based image classification demo (several exist that run entirely client-side, similar in spirit to the TensorFlow.js-based AI Playground elsewhere in AIQ) and have them upload a handful of unusual photos — an object photographed upside down, from very close, or in unusual lighting — and record the model's confidence score on each. The goal is to make the "training data shapes accuracy" idea tangible: unusual photos tend to produce lower confidence or outright wrong guesses, which is a direct, hands-on demonstration of everything covered in Scene 3.