Teaching "Generative AI" to Hacker mode (ages 11–14)

Hook & Warm-Up

Hacker mode drops the mascot and the softer framing — open like you're setting up a real technical investigation, not a story time. Read the app's hook line directly:

"Generative AI doesn't copy — it learns patterns from millions of examples and creates something new. Let's see how AI writes text, makes images, and more."

Push past the hook before opening the app: "If it's not copying, and it's not truly imagining like a person, what IS it doing mathematically? I want a guess at an actual mechanism, not just 'it's smart.'" Take two or three guesses and don't correct them yet — most classes land somewhere near "it's predicting" without being able to say what's being predicted or how. Tell them the lesson is going to name three genuinely different mechanisms — one for text, and two competing ones for images — and that by the end they should be able to tell all three apart, not just describe "AI" as one blob of technology.

Frame the stakes before diving in: "Every earlier lesson in this world was about AI that decides between things that already exist — is this spam, is this a cat. Today's mechanisms all produce something that has never existed before, one piece at a time. That's a genuinely different kind of computation, and it's worth understanding precisely, because these are the exact systems behind the AI tools showing up in the news right now."

Main Activity

Work through the app's three scenes, but push each one to the actual mechanism named in this age band's quiz — that's the standard to teach to.

Scene 1 — How It Creates ✨:

• Next Token (text): This is autoregressive generation — producing output one element at a time, with each new element conditioned on everything generated so far. A model like GPT was trained on huge amounts of text to predict, given a sequence of words (technically tokens — word pieces), the most likely next one. At generation time it does that repeatedly, feeding its own output back in as new input for the next prediction. Ask: "If the model only ever predicts one token ahead, how can it produce something that stays coherent over an entire essay?" (Because each new prediction is conditioned on the whole sequence so far, including everything it just generated — so early sentences constrain later ones, even though there's no separate "planning" step.)
• Diffusion (images): Diffusion models generate images by iteratively denoising random noise, guided by learned score functions. In training, real images are progressively corrupted with random noise until they're pure static, and a network learns to predict, at any noise level, which direction would make the image slightly more realistic — that's the "score function." At generation time, this runs in reverse: start from pure noise and repeatedly nudge it in that learned direction, many small steps, guided the whole way by the text prompt. Explicitly contrast this with the "next token" mechanism: text generation commits to output left-to-right, once; diffusion generates the whole image at once, then repeatedly refines all of it together.
• GANs: The lesson content doesn't name this scene explicitly, but the quiz does, so introduce it here: a Generative Adversarial Network trains two neural networks against each other — a generator that tries to produce convincing fake images from random input, and a discriminator that tries to correctly tell real training images apart from the generator's fakes. Both improve through the competition: as the discriminator gets better at catching fakes, the generator is pushed to make better fakes to fool it. GANs were the dominant image-generation approach before diffusion models became more popular — mention this timeline if students ask "so which one do people actually use now?"
• Video Gen: Video generation extends image generation across time — each frame is produced the way a still image would be, but keeping the frames visually consistent with each other (so an object doesn't randomly flicker or change between frames) is one of the genuinely hardest open problems in the field right now.

Pause here and have students summarize the three mechanisms back to you in one sentence each, in their own words, before moving on — this age band can genuinely confuse "autoregressive" and "diffusion" if the distinction isn't checked immediately: one commits to output left-to-right and never revisits it, the other produces a rough whole image and repeatedly refines all of it together.

Scene 2 — Strengths 💪 and Scene 3 — Limitations ⚠️: Move through strengths quickly (speed, unusual concept combinations, low barrier to entry, cheap iteration) — note that "cheap iteration" is itself a direct consequence of the mechanism: because generation is just running the same trained network again with a different random noise seed or sampling path, producing variation #200 costs the same as producing variation #1, unlike a human artist redrawing something from scratch. Then spend real time on limitations, connecting each to the mechanism above rather than treating it as a separate fact to memorize. Hallucinations are a direct consequence of the training objective: a next-token predictor is optimized to produce plausible-sounding continuations, not verified-true ones, so confident and false can coexist in the exact same sentence. Copyright: these models are trained on huge scraped datasets that include copyrighted art and writing, generally without the creators' consent — ownership of the output, and whether the training itself was lawful, are both live legal questions with no settled answer yet. Deepfakes use the same diffusion or GAN techniques applied specifically to a real person's face or voice, built to deceive rather than merely to create. No Understanding: none of these mechanisms include anything that corresponds to comprehension — a diffusion model has no concept of "sunset," only patterns of pixel values that correlate with the training captions that included that word.

Discussion

• "A next-token predictor has no separate fact-checking step. Given that, is hallucination a bug that could someday be fully fixed, or is it a built-in consequence of how these models are trained? Defend your answer."
• "GANs and diffusion models both generate images, through very different mechanisms. What's an advantage you'd guess one has over the other, based on how each one works?"
• "If a diffusion model is trained on millions of real photographs and artworks without the creators' permission, is it meaningfully different from a human artist who learned by studying other artists' work? What's the same, and what's actually different?"
• "Deepfakes and 'ordinary' AI-generated images use overlapping technology. What specifically — not the technology, but something else — makes a deepfake unethical when a fictional AI-generated portrait isn't?"
• "Temperature controls how random a text model's word choices are. Why might a company set temperature very low for a customer-service chatbot, but higher for a creative-writing tool?"
• "Copyright law hasn't caught up to generative AI yet. If you were writing the law, would you say training an AI on copyrighted art without permission is more like a human artist studying other art, or more like copying? What test would you use to decide?"

Quiz Walkthrough

Wrap-Up & Extension

Close with: "You now know that 'generative AI' isn't one technology — it's at least three different mechanisms wearing the same trench coat: next-token prediction, denoising diffusion, and adversarial generation. Every one of them is a statistical pattern-matcher with zero understanding built in, which is exactly why hallucination, copyright disputes, and deepfakes aren't separate problems — they're the same root cause showing up in text, in ownership law, and in trust."

Extension activity — Mechanism Match-Up: Give small groups four to six real AI tools they may have heard of (for example: ChatGPT, DALL-E or Midjourney, a music generator like Suno, a video generator, a deepfake app, an AI customer-service bot) and have them argue, using only today's vocabulary, which underlying mechanism (autoregressive generation, diffusion, or GAN) each one most likely uses, and what evidence in the tool's behavior supports their guess (for example: "it types out an answer word by word onscreen, which looks like autoregressive generation in action"). Groups present their reasoning, not just their answer — the goal is applying the mechanisms to unfamiliar tools, not recalling a memorized list. This extends the 10–15 minute core lesson into a full period and previews the kind of technical reasoning the Architect-level version of this lesson goes even deeper on.