AIQ
AIQ
Part of the How Computers Read lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 💻 Hacker (11–14) · ⚡ Architect (15–18)
Ask a student to read a sentence off the board out loud. Say "easy, right?" Then read the app's opening line to the class:
"You're reading this right now. Easy, right? For a computer, reading is incredibly hard. It doesn't understand words — it does math on them!"
Ask: "What do you think 'doing math on words' could even mean? How do you do math on a word like 'happy'?" Take a couple of guesses without correcting them — most students will guess something close to "it turns them into numbers," which is exactly right, and you can confirm that: "You're right — that's step one. Today we'll see how a computer turns your words into numbers, and why that lets it predict text, translate languages, and even chat with you."
Follow up with a quick prediction question before opening the app: "If I gave you the word 'happy,' and asked you to guess whether a computer's number for 'happy' would be closer to the number for 'joyful' or the number for 'rock,' which would you guess, and why?" Let a few students explain their reasoning out loud — most will correctly guess "joyful," even without knowing the mechanism yet, which sets up today's big reveal nicely: their intuition about "similar meaning" is exactly what the numbers are built to capture.
One more warm-up prompt worth using if you have a couple of extra minutes: ask "Has anyone's phone or tablet ever finished a word or sentence for you before you were done typing?" Almost every hand will go up. Say: "By the end of today, you'll know exactly how that trick works — and it's the same trick behind spam filters, translation apps, and chatbots like ChatGPT."
Walk through the lesson's three scenes as a guided discussion, pausing to ask questions before revealing each fact.
Scene 1 — Words as Numbers 🔤. Introduce tokens: "A computer breaks a word into smaller pieces called tokens — usually parts of a word. 'Playing' might become 'play' plus 'ing.' Why do you think it splits words up like that instead of keeping them whole?" (Guide toward: fewer total pieces to keep track of, and shared pieces like '-ing' show up in tons of different words.) Write another example on the board and have the class try tokenizing it themselves — "unhappiness" could split into "un" + "happy" + "ness" — and point out that all three of those pieces show up in dozens of other words too, which is exactly why breaking things down this way is efficient.
Then introduce embeddings: "Each token gets turned into a long list of numbers — think of it like GPS coordinates, but for meaning instead of location. Words with similar meanings get coordinates that are close together, so 'happy' and 'joyful' end up near each other on this 'meaning map.'" Reveal the famous example: "King minus Man plus Woman equals Queen — that's real math AI can do on these number-lists, and it usually lands close to the right answer!" Let that land as genuinely impressive; ask, "what does that tell you about what these numbers are actually capturing?" (Push toward: something like the idea of "royalty" and something like the idea of "gender" both got encoded into the numbers, purely from patterns in text, without anyone programming those concepts by hand.)
Scene 2 — Understanding Text 💬. Cover each item as a quick real-world tie-in: Sentiment — "AI can guess if a review is positive or negative just from the words used — 'amazing' leans positive, 'terrible' leans negative." Ask the class to think of one more word that leans strongly positive and one that leans strongly negative. Spam Detection — "Gmail checks the number-patterns in an email against billions of examples it's seen before to decide: real mail, or spam?" Search — "Google doesn't just match your exact typed words — it understands roughly what you MEANT, even with typos." Try it live if you have a projector: search a deliberately misspelled query and show that it still returns sensible results. Autocomplete — "Your keyboard predicts the next word by recognizing patterns from billions of real sentences other people have typed."
Scene 3 — Language AI 🌐. Translation — "Google Translate handles over 100 languages by learning number-patterns that connect millions of already-translated texts." Summary — "AI can shrink a long article into a few sentences by finding the most important patterns in it." Chatbots — "ChatGPT works by predicting the next word based on everything said so far — and it does that prediction over and over, billions of times, across a whole conversation." Q&A — "An AI assistant answers your question by matching the number-pattern of your question to the number-patterns of likely answers." Pause here to connect all four back to the same root idea: "Notice that translation, summarizing, chatting, and answering questions all sound like completely different jobs — but every single one of them is really just 'compare numbers' or 'predict the next number,' over and over."
A good pause point after Scene 1: ask "if a totally new word gets invented tomorrow — say, a brand-new slang word nobody's used before — how do you think a computer would come up with a number for it?" Guide toward: it can often build a rough number just from the sub-word tokens the new word is made of, plus how it's used in the sentences it appears in, even without ever having seen that exact word during training.
Run the practice round as a quick team sort: 📧 Spam filter (AI), 📖 Paper book (not AI), 🌍 Auto-translate (AI), ✏️ Pencil writing (not AI), 📝 Autocomplete (AI), 📰 Printed newspaper (not AI), 🤖 Chatbot (AI), 📨 Postal mail (not AI). For each "not AI" item, ask why — the answer should always be some version of "it does the same fixed thing every time; nothing is predicting or learning."
Before moving on, do a short whole-class recap using the app's own summary line as your script: "Computers can't read like you — they turn words into numbers, then do math to find patterns. Similar words get similar numbers. This is how AI translates, filters spam, and powers chatbots!" Ask one student to restate that in their own words, and one more to add an example that wasn't already covered today (a video app's "recommended for you," a spellchecker underlining a typo, a voice assistant answering a question) — the goal is confirming they can generalize the pattern beyond the exact examples shown.
Close with: "Every time your keyboard finishes a word, an email gets flagged as spam, or a chatbot answers you, remember what's really happening underneath: your words got turned into numbers, and the computer found patterns in those numbers. That's the same trick behind every single example we saw today."
Extension activity — "Guess the Next Word": Write the start of a sentence on the board (e.g., "I opened my lunch box and found a..."). Go around the room and have each student say ONE word they think comes next, based only on what's been said so far — no peeking ahead. After a few rounds, point out: "That's exactly what autocomplete and chatbots do — guess the next word based on everything said before, one word at a time, over and over." For a harder round with older Builders, try a sentence with less obvious context (like "The strangest thing I ever saw was...") and discuss why the guesses vary more from student to student than they did with the lunch box sentence.
Wrap the activity by connecting it back to real embeddings: split the class into pairs and give each pair two word cards (like "happy"/"joyful" for one pair, "happy"/"rock" for another). Ask each pair to rate, on a scale from 1 (totally different) to 10 (basically the same), how close their two words' number-codes should be — then have pairs share their ratings and reasoning. This stretches the 10–15 minute core lesson into a full class period and makes both the "predict the next word" idea and the "similar meanings, similar numbers" idea something they've physically practiced rather than just heard about.