Teaching "AI in Your Pocket" to Hacker mode (ages 11–14)

Hook & Warm-Up

Open with the real hook line, delivered without softening it: "Your phone runs dozens of AI models at once — some on the device, some in the cloud. Let's peek behind the apps you use every day."

Follow with a quick gut-check question: "Pull up your phone's home screen in your head. How many apps do you think are running some kind of AI right now, in the background, even if you're not actively using them?" Take a few guesses out loud — most students will underestimate. "The honest answer is: probably more than half of what's on that screen. Some of it runs right here, on the chip in your hand. Some of it runs on a server somewhere and just sends you the result. Today we're figuring out which is which, and why it matters."

Frame the stakes: "This isn't just trivia. Understanding what's AI and what isn't — and where it runs — is the first step to understanding why your feed looks the way it does, and why your phone knows things about you that you never typed in."

Set the tone for the lesson explicitly, since this age band can handle nuance without needing it softened: "We're not here to scare you off your phone, and we're not here to tell you it's all magic and wonderful either. We're going to look at how these systems actually work, mechanism by mechanism, and then decide for ourselves what we think about them."

Main Activity

Move through the three scenes with more technical precision than younger bands, naming the actual mechanism behind each example, not just the effect.

Scene 1: Phone AI (on-device).

• Face ID — runs a neural network locally on the phone's chip. It projects an infrared dot pattern onto your face, builds a depth map, and compares it against a stored mathematical representation. Ask: "Why would Apple want this running on the device instead of sending your face data to a server?" Guide toward: speed (no network round-trip) and privacy (biometric data never leaves the phone).
• Smart Keyboard — a language model personalized to your typing history, adapting its predictions to your vocabulary rather than a fixed dictionary.
• Battery AI — learns your daily charging pattern and slows charging near 100% overnight to reduce battery wear over time.
• Ruler App — not AI. Fixed arithmetic: pixel count converted to a physical unit using a known screen density. No training, no data, no adaptation.

Preempt a common confusion at this age: "on-device" doesn't mean "simple." Face ID's neural network is a genuinely trained model, just one small and efficient enough to run on a phone chip rather than a data center. The distinction that matters isn't device-vs-server complexity — it's whether the system learned a pattern from data (AI) or was given a fixed procedure by a programmer (not AI), regardless of where it happens to run.

Scene 2: Social & Entertainment (cloud-based recommendation systems). This is the core technical content for this age band — spend the most time here.

• TikTok's algorithm — combines collaborative filtering (what similar users engaged with), content-based signals (analyzing the video itself — audio, visual features, captions), and engagement prediction (will you finish it, rewatch it, or skip it). It's not one technique, it's a stack of them working together, adapting within roughly 40 minutes of viewing.
• Netflix — an estimated 80% of watch time is attributed to its recommendation system, not to active searching. Ask: "What does that number tell you about how much control you actually have over what you end up watching?" A useful follow-up: "If you canceled your account and re-signed up with a totally blank history, would the recommendations start over from nothing, or would the system fall back on something else in the meantime?" (Most such systems fall back on general popularity trends for new or "cold-start" users until enough personal data accumulates — worth naming that recommendation AI isn't magic on day one, either.)
• FM Radio — not AI. Broadcast on a fixed frequency; identical signal to every receiver, no personalization layer at all.
• Spotify Discover Weekly — analyzes patterns across roughly 4 billion playlists to surface songs statistically similar to your taste that you haven't heard yet.

Introduce the filter bubble concept here directly: "A system optimized to show you more of what you already engage with can, over time, narrow the range of content and viewpoints you actually see — even if no single recommendation is 'wrong.'"

Scene 3: Shopping & Maps.

• Shopping recommendations — a large share of purchases on major platforms are attributed to "customers also bought"-style AI recommendations.
• Google Maps — predicts ETA by combining live traffic data, historical traffic patterns for that road and time, and known events (concerts, road closures) — not just current speed.
• Paper List — not AI. No data, no model, no adaptation.
• Visual Search (Google Lens) — this is computer vision: an object-recognition model identifies what's in the camera frame and matches it to a product database.

Worth a quick aside on Google Maps: predicting a traffic jam before it happens requires two different kinds of data working together — real-time signals (phones currently moving slowly on that road) and historical patterns (this road is always slow at 5pm on a Friday), plus a layer of known events. That's a small forecasting system in its own right, not a single lookup.

Bring in computational photography as a closing example: modern phone cameras don't take one photo, they capture multiple frames at different exposures and use AI to fuse them (multi-frame processing, scene detection, HDR fusion) — which is why phone photos look far better than the raw sensor hardware alone would produce.

Wrap the walkthrough with a quick synthesis question before moving to discussion: "Looking back across all three scenes, which examples ran locally on the phone, and which ones needed a server somewhere? What's the pattern?" Students should be able to reason toward: biometric and personal-habit models (Face ID, keyboard, battery) tend to stay on-device for speed and privacy; large-scale recommendation and search systems (TikTok, Netflix, visual search) need a server because they're drawing on data from millions of users and updating constantly, which no single phone could store or compute alone.

Discussion

These work well as a think-pair-share: give students 60 seconds to think alone, a minute to compare with a partner, then open to the full class. This age band tends to give more thoughtful answers when they've had a moment to form an opinion before being asked to share it publicly.

• "What's the practical difference between an AI model that runs on your phone versus one that runs on a company's server? Think about speed, privacy, and how often it can update." — Push toward all three axes: on-device is faster (no network round-trip) and keeps sensitive data local, but a server-side model can be retrained continuously on data from millions of users, which a single phone never could.
• "If TikTok's algorithm adapts to you within 40 minutes, what does that suggest about how much of your feed reflects a deliberate choice on your part versus a fast feedback loop you didn't fully choose?" — There's no single right answer, but the honest tension is: your early choices matter a lot, and they get amplified quickly, in a way that's easy to not notice happening.
• "Is a filter bubble something a company chooses to build, or a side effect of any system built to maximize engagement? Does that distinction matter?" — Both can be true at once: the effect can be an unintended side effect of the optimization target, and a company can still be responsible for choosing that target in the first place.
• "Computational photography means the 'photo' your phone gives you is already processed and combined by AI before you ever see it. Does that change how you think about a photo as 'real'?" — Encourage students to distinguish "manipulated to deceive" from "processed to look good" — computational photography is closer to the second, but it's worth naming that the line between the two isn't always obvious.
• "Amazon and Netflix report large percentages of purchases/watch time coming from recommendations. Who benefits most from those systems being that effective — you, or the company?" — A fair answer is usually "both, but not necessarily equally" — you save time and often find things you genuinely like, while the company's revenue and engagement metrics both improve regardless of whether the recommendation was actually in your best interest.

Quiz Walkthrough

Wrap-Up & Extension

Close with: "Every scroll, tap, and pause you make on your phone is a data point feeding a system that's constantly re-calculating what to show you next. Knowing that doesn't mean you have to quit your apps — it means you get to use them with your eyes open instead of on autopilot."

Extension activity (20–25 minutes): Have students open one recommendation-driven app (with a family member's permission if needed) and, for five minutes, deliberately engage with content on a topic they're not actually interested in — fully watching several videos on it. Have them check back the next class period (or later that day) to see whether their feed shifted toward that topic, and write a short paragraph on what that experiment reveals about how quickly and mechanically the recommendation system responds to behavior, regardless of genuine interest.

Alternative or additional extension (if devices at home aren't reliable for every student): run a paper-based version in class. Give small groups a fictional "user's" last ten watched videos (write a simple list — five cooking videos, three comedy skits, two skateboarding clips) and have them predict what the recommendation system would show next, and why. Then reveal a twist: the user watched one video by accident and skipped it immediately. Ask groups to reconsider their prediction — this makes the mechanical, behavior-driven nature of the system concrete without needing an actual app open.

Timing note: the three-scene walkthrough runs roughly 15–18 minutes at this depth, discussion 8–10 minutes, and the quiz 5–6 minutes — a full core lesson of about 30 minutes, with either extension comfortably filling the remainder of a standard period.

Send students out with a genuinely open question rather than a tidy conclusion: "None of what we covered today means these apps are bad — it means they're built, on purpose, by people making specific choices. The next time your feed feels a little too accurate, that's not a coincidence you should find creepy — it's a system working exactly as designed. What you do with that knowledge is up to you."

If your class continues into later AIQ lessons on fairness and bias in AI, flag the connection now: the same mechanism that personalizes a feed usefully can also, unchecked, personalize discrimination — a system that's just "learning patterns from data" reflects whatever patterns and biases exist in that data, for better and for worse. This lesson is the mechanical foundation that later ethics-focused lessons build on.