AIQ
AIQ
Part of the AI or Not? The Game lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · 💻 Hacker (11–14)
Architect-mode students are old enough to handle — and to benefit from — the genuinely unresolved edges of this question, so open by naming the boundary as contested rather than settled:
"The boundary between AI and non-AI is less clear than most people think. Let's examine edge cases, from rule engines that mimic intelligence to hidden ML models in everyday products that most users never notice."
Start with a short framing lecture rather than a game, since this age group is being prepared to evaluate real claims, not just play a sorting game. Introduce the core distinction precisely:
"The AI/non-AI boundary depends on adaptability. Expert systems with hand-crafted rules — even very complex ones — aren't machine learning. Statistical models that update their internal parameters through training on data are. Modern products often embed ML in unexpected places: smartphone battery optimization, email 'smart reply' features, even health-analytics toilets."
Push the class to articulate why this distinction matters beyond trivia. Ask directly: "If you were evaluating a vendor's product for your company, or reading a startup's pitch deck, why would you care whether their 'AI feature' is actually machine learning versus a well-disguised rule engine?" Let a few students attempt an answer before the activity — you'll return to this question explicitly in the discussion section, once they've seen concrete cases.
It's worth being upfront with this age group about why the field itself is inconsistent on this question. "Artificial intelligence" as a research discipline has, since the 1950s, included symbolic reasoning, search, planning, robotics, and statistical learning — machine learning is the sub-field responsible for the current wave of consumer products, and it's the one this lesson's "does it learn from data" test is really probing. Being precise about this distinction — AI as the broad field versus ML as the specific data-driven technique — is itself a marker of genuine technical literacy that separates a careful reader of tech claims from a casual one.
Move through the app's three rounds as source material for a technical discussion, treating each item as a short case study rather than a simple label to assign.
Calculator vs. auto-DJ: a calculator's arithmetic is deterministic and stateless with respect to usage history — the function mapping input to output never changes based on prior interactions. An auto-DJ feature, such as Spotify's, is built on models trained on aggregate listening data (skips, replays, co-occurrence of tracks in playlists) that infer a specific user's likely preferences and select or sequence tracks accordingly; its output distribution for a given user shifts as more behavioral data accumulates. That shift — parameters or predictions changing in response to data, rather than staying fixed — is the operative distinction.
Digital clock vs. real-time voice-changer: a digital clock's behavior is governed by a fixed physical process (crystal oscillation) counted by fixed logic — there is no data-driven adaptation anywhere in the pipeline. A convincing real-time voice transformation typically requires a model that has learned a mapping between acoustic features (pitch, formants, timbre) across voices, trained on large corpora of speech — the "real-time" constraint additionally requires that inference be fast enough to run with imperceptible latency, which is itself a nontrivial systems-engineering problem layered on top of the modeling problem.
Spend the most time here, since it anchors two quiz items. Gmail's mail classification (spam filtering, and separately, "Smart Reply") is worth treating as two distinct systems. Spam classification is a supervised learning problem: models are trained on large labeled datasets of spam versus legitimate mail and continuously refined as new patterns emerge. Smart Reply, historically, was documented by Google's research team as built on sequence-to-sequence neural networks — models that read an entire input sequence (an email's text) and generate an output sequence (a short candidate reply) — the class of architecture is worth naming explicitly here because it's the correct answer the quiz is testing, and it's a genuinely different technique from the classification approach used for spam filtering.
Battery management is the other anchor. Frame it carefully: some phone battery-optimization systems, and a meaningful line of research on the problem, model power management as a reinforcement learning problem — an agent (the optimization system) learns a policy for allocating power and scheduling background activity by receiving reward signals (e.g., battery life extended, user-perceived performance maintained) through repeated trial across a device's actual usage. This is a different learning paradigm from the supervised classification used in spam filtering: no fixed labeled "correct answer" exists for the ideal power policy — the system has to discover good behavior through interaction and feedback over time.
Use this pairing to make an explicit point about paradigm selection: the choice between supervised learning and reinforcement learning isn't arbitrary, it follows from the shape of the problem. Spam filtering has abundant ground-truth labels (millions of emails users have already marked as spam or not), which is exactly what supervised learning needs. Optimal battery policy has no such ground truth — there's no dataset of "correct" power decisions waiting to be labeled — so a paradigm built around trial, feedback, and reward is the natural fit instead. Ask students to name one more real-world problem for each paradigm and defend the fit.
Precision agriculture is a clean computer-vision case study: convolutional models trained on labeled aerial or satellite imagery (healthy vs. stressed vs. diseased vegetation, often cross-referenced with spectral bands beyond visible light) are deployed to flag problem regions in a field faster and earlier than manual inspection would catch them. Health-analytics toilets are a legitimate, if niche, applied-sensing category — worth using as a springboard into the discussion question about differential stakes: a misclassification in a music recommender costs a bad song suggestion, while a misclassification in a health-monitoring device carries a materially different risk profile, which is precisely the kind of distinction a competent AI evaluator needs to reason about.
Close the guided pass with the stapler as the clean non-AI endpoint and AI-generated art as the clean AI endpoint (a generative model trained on large corpora of images or text, sampling novel output conditioned on a prompt), then have students complete the app's sorting round independently, explicitly asking them to be ready to defend their classification of at least one item they found ambiguous.
Close by tying the lesson explicitly to a career-relevant skill: "Whatever field you go into — product management, journalism, policy, engineering, medicine — you will be handed 'AI-powered' claims that need evaluating. The test isn't complicated: does the system's behavior change because it processed data, or because a person edited its rules? Everything else — how impressive it sounds, how much it costs, how it's marketed — is noise relative to that one question."
Extension activity — AI Claim Due-Diligence Memo (25–30 minutes, portfolio-eligible): Have each student choose one real product or feature that markets itself as "AI-powered" (a school tool, a consumer app, something from current tech news) and write a one-page due-diligence memo, as if advising a decision-maker on whether to trust or adopt it. The memo should: (1) state what evidence would distinguish genuine machine learning from rule-based automation in this case, (2) identify what data the system would need to collect to function as claimed, and (3) flag one concrete risk specific to that data collection or to the system's failure mode. This is a strong candidate for the Architect portfolio — it demonstrates applied critical evaluation of an AI claim, which is a skill directly transferable to nearly every career path this age band is considering.
For students on a technical career track, offer an optional harder variant: instead of a product already on the market, have them sketch which learning paradigm (supervised, unsupervised, or reinforcement learning) would be the best technical fit for a hypothetical AI feature of their own choosing, and justify the choice using the same reasoning applied to the spam-filter/battery-management contrast earlier in the lesson. This pushes the exercise from "evaluate someone else's claim" to "design a defensible one," which is the skill a technical career path in this space will eventually demand.