AIQ
AIQ
Part of the When AI Gets Confused lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 💻 Hacker (11–14) · ⚡ Architect (15–18)
Open with a quick show of hands: "Who here has ever used a translation app, a photo filter, or asked a voice assistant a question, and gotten a weird or wrong answer back?" Let two or three kids describe what happened — a mistranslated word, a filter that put ears on the wrong spot, Siri or Google Assistant misunderstanding a question. Say: "Today we're going to find out exactly why AI does that — and it's not because it's broken." Then read the lesson's hook line together, projecting it if you can:
"AI seems smart, but it can be fooled surprisingly easily!
A tiny sticker on a stop sign can trick a self-driving car. Let's explore AI's weak spots."
Ask: "How many stickers do you think it would take to trick a self-driving car's stop-sign reader — a hundred? Ten?" Take a few guesses, then reveal: "Researchers did this for real, and it took just a handful of small stickers, placed in exactly the right spots, to make an AI camera read a stop sign as a completely different sign." Let that land for a second — this is the moment that hooks Builder-age kids, because it's a real, verifiable claim about something they know matters (traffic safety), not a cartoon example.
Mention the badge on the way in, since Builder mode runs on XP and badges rather than stars: "This lesson earns the 'Glitch Finder' badge — you're literally being trained to spot AI's weak spots, like a bug hunter." That framing (finder, not victim) sets the tone for the whole lesson: the goal isn't to make AI look bad, it's to build the specific skill of noticing when to double-check it.
Work through the app's three scenes as a guided tour, pausing after each one to check understanding with a quick "why did that happen?" question rather than just reading the facts aloud. Builder-age students can handle real cause-and-effect, not just "isn't that funny" — so for each card, push one level past the fact itself to the reason behind it. That habit is what separates this age band's version of the lesson from Explorer's: same examples, but with an actual explanation attached to each one instead of just a reaction.
Scene 1 — Vision Fails 👁️. Show the Stop Sign card (🛑) and explain the sticker trick, then ask: "If a person walked up to that same sign, would they be fooled?" (No — a person understands what a stop sign means, not just what it looks like.) Show the leopard-sofa card (🐆): "AI once looked at a leopard-print couch cushion and confidently called it a leopard." Ask: "What do you think the AI was actually looking at — the whole picture, or just one part of it?" (Just the spotted pattern.) Show the banana card (🍌): "Someone added a tiny, invisible change to a photo of a banana, and it made an AI think the photo was a toaster — a change so small a human couldn't even see it." Show the glasses card (😎): "Special patterned glasses can trick facial recognition into ID'ing you as a totally different person." Tie it together: "In every one of these, the AI wasn't looking at the whole picture the way you do — it was matching a narrow pattern, and something in the pattern got hijacked." If a student asks "so is this basically hacking a computer?" that's a great connection to make explicit: "Sort of — it's called an adversarial example, and it's a real area of AI safety research, where people deliberately try to find an AI's blind spots before someone with bad intentions does."
Scene 2 — Language Fails 💬. Read the sarcasm example out loud in a flat, sarcastic tone: "Oh great, another Monday." Ask: "What did I actually mean? How did you know?" (Tone of voice, facial expression, context — none of which text alone carries well.) Cover the "bank" example (🏦): a river bank versus a money bank — ask students for another English word with two meanings (spring, bat, bark are good ones) to show this isn't a one-off. Cover the nonsense card (🤪): AI can write a whole confident paragraph about "why rocks make good pets" — ask a student to improvise one sentence of that paragraph out loud, for a laugh, then ask "was that convincing, even though it's obviously silly?" Finish with made-up facts (📅): AI can invent a date, quote, or statistic that sounds exactly as certain as a real one.
Scene 3 — Why AI Fails 🔍. This is the "so what" scene — spend real time here, since it's where the earlier examples turn into an actual explanation rather than a list of funny glitches. Data Limits (📊): "AI only knows what was in its training data — if it never saw a purple banana, it has no good answer for one." Ask: "So if you wanted an AI to recognize something really unusual, what would you need to give it?" (Lots of examples of that unusual thing.) No Common Sense (🧩): "You know a car can't fly without being told — AI doesn't have that kind of built-in knowledge; it only has patterns." Context Blind (🌍): "AI can't always tell if a photo of a toy gun is dangerous or harmless — that judgment needs context a picture alone doesn't carry." Defenses (🛡️): "Engineers test AI with tricky examples on purpose — sometimes called 'stress-testing' — and build in safety layers, but no AI is 100% foolproof, which is exactly why a human should stay involved." Close the scene by connecting all four cards back to one sentence: "Every single failure today traces back to the same root cause — AI matches patterns from its training data, and patterns are not the same thing as understanding."
Optional quick pairs activity: If time allows, split the class into pairs and have them flip through the app's "practice" cards from this lesson (adversarial sticker attack, sarcasm detection fail, AI hallucination, confident nonsense, versus their human counterparts) and sort each one out loud: "is this an AI weak spot, or something humans naturally handle better?" This is a fast, low-stakes way to check that the sorting logic — not just the individual facts — has landed before moving to discussion.
These work well as a think-pair-share: give thirty seconds of silent thinking, sixty seconds to compare answers with a partner, then open it up to the whole class. That structure gets more kids actually reasoning through an answer instead of just waiting for a confident classmate to speak first.
Builder mode's version of this quiz uses the same four questions as the base lesson but shuffles the answer order, so don't rely on remembering "which letter" was correct in a preview — read the options fresh with the class. Each question maps directly back to one of the three scenes above, which is worth pointing out explicitly: "this next question is about the stop sign scene we just did."
Close with: "Today you learned that AI doesn't fail randomly — it fails in specific, understandable ways, because of how it learns. That means you can get better at spotting when to trust it and when to double-check, which is exactly the kind of thinking that makes you smarter than the tool you're using."
Extension — "Trick the AI" design challenge: In pairs, have students design (on paper, no coding needed) one example of something that could confuse an AI, choosing from a vision trick, a language trick, or a "why it fails" scenario like the ones covered in class. Each pair writes: (1) what the trick is, (2) what a human would correctly see or understand, and (3) why the AI might get it wrong instead. Give them eight to ten minutes, then have two or three pairs present their trick to the class as a mini "case file." Encourage them to riff on a local, familiar example rather than reusing one from the lesson — a jeepney sign partly covered by a vendor's umbrella, a Taglish sentence mixing English and Filipino words in a way that could confuse a translation app, or a school ID photo taken at a strange angle are all fair game and often lead to better discussion than the app's own examples. This turns the passive examples from the lesson into something students construct themselves, which is a stronger test of whether they actually grasped the "pattern-matching versus understanding" idea than simply recalling the four scenes.