AIQ
AIQ
Part of the AI in Science lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · ⚡ Architect (15–18)
Open with the framing this age band is ready for — AI as a working tool inside real research, not a gimmick. Say: "AI is speeding up science — finding planets, predicting weather, and mapping proteins. Let's see how AI is becoming every scientist's research partner."
Warm-up question: "Think of a scientific field you're interested in — could be space, medicine, climate, whatever. What's a problem in that field that's mostly about finding a needle in a haystack of data?" Give them a minute to think, maybe in pairs, before taking two or three answers. You're priming them to notice that a huge amount of scientific work is fundamentally a search problem — which is exactly the kind of problem AI is good at.
Bridge in: "Today's examples are all real, published science — a lot of it from just the last few years. As we go, I want you thinking about a specific question: in each example, is the AI actually doing science, or is it doing something else that helps scientists do science?" That framing sets up the discussion and quiz questions about correlation versus causal understanding later in the lesson, so plant it early.
If your class has covered the machine learning basics earlier in AIQ (how a model learns from labeled training examples), briefly reconnect it here: "Everything we're about to look at is the same core idea from those earlier lessons — a model trained on a huge number of past examples — just applied to space telescopes, weather satellites, and protein databases instead of photos or text. The domain changes; the underlying method mostly doesn't." This keeps the lesson from feeling like a disconnected set of trivia and ties it back to the curriculum's throughline.
Hacker mode students can handle more mechanism, not just more vocabulary. For each scene, go one level deeper than "AI found a pattern" — explain roughly *how* the pattern-finding works, and use the built-in "no AI needed" comparison in each scene to sharpen the distinction between AI and ordinary tools.
Scene 1 — Space & Physics. Present Planet Hunter, Dark Matter, Stargazing, and Black Hole Image. Explain the transit method concretely: "When a planet passes in front of its star from our point of view, the star's measured brightness drops by a tiny, precise amount, then recovers. NASA's Kepler mission recorded years of brightness data from hundreds of thousands of stars, and AI models were trained on confirmed examples of that dip-and-recover shape to recognize it automatically in new data — something that would take human astronomers an impractical amount of time to check star by star." For dark matter: "AI-powered simulations model how billions of virtual particles would behave under different physical assumptions, then compare the simulated result to what real telescopes observe — that comparison is how scientists test theories they can't directly experiment on." Ask: "Why can't scientists just run a real experiment on dark matter the way you'd run one in a chemistry lab?" (Good answer: dark matter can't be directly manipulated or observed — simulation is one of the only ways to test theories about it.)
For the black hole image, add the layer of nuance this age can handle: "The 2019 image wasn't a photograph in the normal sense — it was reconstructed computationally from radio data collected by telescopes on different continents, all pointed at the same target at the same time. A later 2023 reprocessing of that same data, called PRIMO, specifically used machine learning trained on simulated black holes to sharpen the reconstruction." Ask: "What's the difference between 'AI took a photo' and 'AI helped reconstruct an image from raw data'? Does that distinction matter?"
Scene 2 — Climate & Earth. Present Ocean Monitoring, Weather AI, Growing Plants, and Wildfire AI. Explain: "AI models trained on labeled satellite imagery can scan huge volumes of images to flag coral bleaching, temperature anomalies, or plastic pollution far faster than a team of scientists reviewing images by hand. For weather, models like Google DeepMind's GraphCast are trained on decades of historical atmospheric data and learn to predict how weather patterns tend to evolve — which is fundamentally different from how traditional weather forecasting works." This is the moment for the key distinction: "Traditional weather models solve physics equations describing how air, heat, and moisture move — that's called numerical weather prediction. AI weather models instead learn statistical patterns from historical data. GraphCast can produce a 10-day forecast in under a minute, versus hours on a supercomputer for the physics-based approach — and on many measures, it's more accurate." Ask: "What's a risk of a model that's really good at recognizing 'weather tends to look like this' from history, but isn't directly solving the physics?" (Good answer: it might struggle with genuinely unprecedented conditions that don't resemble its training data.)
Scene 3 — Biology & Chemistry. Present Protein Folding, Virus Tracking, Basic Microscope, and Material Discovery. Explain: "A protein's function depends on its 3D shape, which is determined by how its chain of amino acids folds. Figuring out that shape experimentally, protein by protein, was slow and expensive — decades of work. AlphaFold was trained on the relatively small set of protein shapes already solved in labs, and learned to predict a new protein's shape directly from its amino acid sequence. It's now produced predicted structures for over 200 million proteins — essentially the entire catalog known to science — which the field genuinely considers a 50-year grand challenge solved." For materials: "A similar approach let AI propose 380,000 new stable material candidates by learning from known crystal structures — but every one of those still needs to be synthesized and tested in a real lab before it's a usable material, not just a prediction." Ask: "If AlphaFold is a prediction, not a lab measurement, why does the scientific community still treat it as a genuine breakthrough?" (Good answer: its predictions are validated against real experimental data at high accuracy, and it lets scientists prioritize which proteins are worth the time and cost of lab confirmation.)
Also mention Virus Tracking, especially with a class old enough to remember COVID-19 directly: "AI helped track how the virus spread globally by analyzing case and mobility data, and helped accelerate vaccine design by working with the underlying protein structures involved." This is a good moment to connect back to the medical AI content earlier in World 4 if your class has already covered it: "How is this similar to, or different from, the diagnostic AI examples we saw in the healthcare lesson?" (Good connection: both rely on pattern recognition in large datasets and both still require human experts — clinicians or virologists — to validate and act on the results.)
Close the activity by returning to the framing question from the hook: "So — in each of these, is the AI doing science, or helping scientists do science faster? Where's the line?" Push the class toward the idea that AI in these examples generates hypotheses, narrows search spaces, and makes predictions — but the underlying scientific method (form a hypothesis, test it, validate against reality) still requires human-designed experiments and human interpretation.
Timing note: with the deeper mechanism explanations and discussion prompts woven in, this activity runs 15–18 minutes. If time is tight, the black hole nuance and the numerical-vs-AI weather distinction are the two richest discussion moments — protect those over the materials discovery example if you need to cut something.
These questions ask students to reason about *why* something is true, not just recall a fact — walk through the reasoning for each answer rather than just confirming the correct letter.
Close with something like: "AI is becoming a real research partner across science — generating hypotheses, running simulations, and finding patterns in datasets too big for any human team to search by hand. But every example we saw today still depends on human scientists to ask the right questions, design real experiments, and figure out what a result actually means. That partnership, not replacement, is the pattern to watch for."
Extension activity — "Correlation vs. Causation Hunt." In small groups, have students find (or you provide) two or three real or invented examples of a strong statistical correlation that isn't a causal relationship — classic examples include ice cream sales correlating with drowning rates (both actually driven by hot weather), or countries' chocolate consumption correlating with Nobel Prizes won. Have each group explain what the real underlying cause is likely to be. Then connect it back: "This is exactly the trap an AI model trained purely on correlations can fall into if a scientist isn't careful about interpreting its output." This stretches the core lesson into a full 25–30 minute period and gives concrete practice with the lesson's most sophisticated idea.
For a class that finishes early or wants an extra challenge, add a follow-up round: have each group invent their own scientific-sounding but fake correlation (for example, "years since a country's first satellite launch correlates with life expectancy") and challenge another group to guess whether it's a real published correlation or a made-up one, and to propose what the real underlying cause might be if it does turn out to be a genuine correlation. This turns the lesson's central caution — that AI can find real patterns without any causal explanation — into something students construct and defend themselves, which tends to stick longer than a caution they're just told to remember.