AIQ
AIQ
Part of the AI in Healthcare lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · ⚡ Architect (15–18)
This age group can handle — and often enjoys — a more direct, slightly clinical framing. Don't oversell it; let the real content carry the interest.
Say: "AI is changing medicine. It spots diseases in scans. It discovers new drugs. It personalizes treatment. Let's see how machine learning helps doctors save lives."
Follow with a pointed question: "Here's a claim you'll see in tech headlines a lot: 'AI is more accurate than doctors.' Is that true, false, or more complicated than that?" Let a few students argue it out — don't resolve it yet. Tell them: "By the end of this lesson you should be able to give a better answer than a headline can." That sets up the quiz's question on FDA classification and the limitation question as the payoff.
Have students work through the lesson's three scenes individually or in pairs, then debrief each one with the discussion prompts below — at this age, the goal is connecting each example to a mechanism, not just a fact.
Scene 1 — Diagnosis. The lesson covers three AI diagnostic tools: an X-ray model that detects lung cancer with accuracy the lesson describes as sometimes matching radiologists; a Google-built model that screens retina photos for diabetic blindness, aimed at regions with limited access to eye specialists; and a skin-cancer classifier trained on 130,000+ labeled images that performs comparably to dermatologists on melanoma detection. Paired against those is an ordinary stethoscope — pure acoustic physics, no learned pattern involved. Debrief: "All three AI examples needed a large set of labeled images to learn from. What do you think 'labeled' means here, and why does the quality of those labels matter?" Push toward: the labels came from real doctors' diagnoses, and if those labels were wrong or inconsistent, the model would learn the wrong pattern.
Worth pausing on for this age group: reported accuracy figures for medical AI models ("sometimes matching radiologists," "comparable to dermatologists") almost always come from a specific test dataset under specific conditions, not a guarantee of real-world performance everywhere. A model can score very well on the images it was tested on and still perform worse on a different hospital's equipment, a different population, or images taken under different conditions. This is exactly why regulatory validation (covered later in this activity) involves testing beyond the original training data before a tool is cleared for real clinical use.
Expect some students to push back with "but if it's 94% accurate, isn't that basically solved?" — a good, common reaction worth addressing directly. Point out that 94% accuracy still means roughly 1 in 17 predictions is wrong, and ask what that error rate would mean applied to, say, every patient in a busy hospital in a single day. This usually reframes "94%" from sounding impressively high to sounding like a number that still requires human review — exactly the conclusion the lesson is building toward.
Scene 2 — Treatment. Covers an AI system credited with helping find a new antibiotic in 2020 by computationally screening 100 million compounds — work that would take a human team decades to do by hand; an AI used in gene therapy to help design cancer treatment personalized to a patient's specific tumor DNA; and AI-assisted surgical robots that give surgeons steadier, magnified control during procedures (the robot does not operate autonomously — a surgeon operates it). Debrief: "A drug-discovery AI narrows 100 million candidates down to a shortlist. What still has to happen after that, before any of those candidates becomes an actual medicine?" Answer: lab synthesis, animal testing, then multiple phases of human clinical trials — the AI accelerates the search phase, not the whole pipeline.
It's worth naming the actual mechanism a bit more precisely for this age group: these models are typically trained to predict a property of a chemical compound — how likely it is to bind to a target molecule, or how likely it is to be toxic — based on patterns learned from a large database of compounds whose properties are already known from lab experiments. That turns "search 100 million compounds" from a brute-force lab problem into a fast computational ranking problem, after which only a small, high-probability shortlist needs real lab testing. The model doesn't know chemistry the way a chemist does — it has learned a statistical association between certain molecular patterns and certain outcomes.
Scene 3 — Patient Care. Covers AI-driven irregular-heartbeat detection in smartwatches, an AI chatbot (Woebot) offering CBT-style techniques for mental health support, and AI that analyzes patterns across millions of health records to flag patients at risk of getting sick. Debrief: "A model trained on millions of health records to 'predict who might get sick' — what's one way that could go wrong, or be unfair, if the records it trained on weren't representative of everyone equally?" This is a direct on-ramp to the quiz's training-data-bias question, so let students reason toward it themselves before the quiz gives the answer away.
Woebot is worth a moment of extra care in this scene, since some students may have real experience with mental health apps or may know someone who does. Keep the framing neutral and factual: it's a real, documented example of AI applied to mental health support, offering structured techniques drawn from cognitive behavioral therapy, available at any hour — genuinely useful for some people as a supplement, and explicitly not positioned, even by its own makers, as a replacement for a licensed therapist in cases of serious mental health need. If the conversation turns personal, redirect gently to the general principle rather than probing individual students' experiences.
It's also worth distinguishing "detection" from "prediction" explicitly here, since students can otherwise blur the two: the smartwatch example detects an irregularity that's already happening (an unusual heart rhythm right now), while the health-records example predicts a future risk based on patterns in someone's history — a meaningfully different, and generally harder and less certain, kind of claim. A flagged future risk is even more clearly a prompt to get checked by a professional, not a diagnosis, precisely because it is a probability about something that hasn't happened yet.
These are meatier questions than the base app quiz covers, and several don't have a single clean answer — that's intentional at this age. Let disagreement stand where it's genuinely a judgment call (the accountability question especially), rather than steering the class toward one "correct" opinion.
Every question in this quiz targets a specific nuance the base-lesson quiz doesn't cover at younger ages — regulatory classification, the mechanism of drug-discovery AI, the correct framing of "decision support," and a genuine, documented limitation. If a student answers correctly but can't explain why, that's worth a follow-up question before moving on, since getting the right answer without the reasoning behind it is exactly the surface-level understanding this age band's activity is meant to go beyond.
Timing note: the in-app content plus scene debriefs and discussion typically take 20–25 minutes for this age group given the depth of the follow-up questions above; the extension activity below adds another 20–25 minutes, so plan for close to a full class period if you run everything. If time is tight, the discussion questions alone are enough to hit the lesson's core learning objectives without the group-research extension.
Close with: "So — 'AI is more accurate than doctors' isn't really a yes-or-no claim. AI can be very accurate at a narrow task it was trained for, on the kind of patient its training data represented well. That's powerful, but it's not the same as replacing a doctor's broader judgment — which is why every deployment of this technology keeps a human in the loop."
Extension activity (20–25 minutes): Split the class into small groups and assign each group one of the four scene examples (X-ray AI, drug discovery, smartwatch heart monitoring, or health-record risk prediction). Each group researches — or reasons through, if internet access isn't available — one question: "What's the worst realistic way this specific tool could go wrong, and what safeguard would prevent it?" Have each group present their failure mode and safeguard in two sentences. This turns the training-data-bias idea from an abstract quiz answer into something students derive for a concrete case themselves.
If your class has more time, a good follow-up is to have each group also identify one Philippine-specific consideration for their assigned tool — for example, whether a smartwatch-based heart monitor is realistically affordable and accessible for most Filipino families, or whether a health-record risk-prediction model trained mostly on data from other countries might behave differently applied to Filipino patients. This grounds an otherwise abstract global-technology conversation in the students' own context, and it's a natural bridge into later AIQ lessons on AI ethics and access.