AIQ
AIQ
Everything a teacher needs to deliver this lesson — pick your grade's script below once you've read the background.
Like every AIQ lesson, this one is a short core loop — a hook, three "scenes" of tappable examples, and a quiz — built for one 5–15 minute sitting. By the end, a student should be able to:
You do not need a science or computer science background to teach this lesson well. Every example in "AI in Science" is a variation on the same idea the rest of AIQ builds on: an AI system gets good at one narrow job by studying an enormous number of past examples of that job done correctly, and then makes a prediction about a new case it has never seen. Here, the "job" just happens to be astronomy, climate science, or biology instead of, say, recognizing a photo.
Space and physics. When a planet passes in front of its star, the star's brightness dips by a tiny, precise amount. Telescopes like NASA's Kepler mission collected years of brightness data from hundreds of thousands of stars, and AI models were trained to recognize that faint dip-and-recover pattern among an ocean of noisy data — a pattern too subtle and too repetitive across too many stars for a team of humans to check by eye. The same idea — finding a signal buried in a mountain of numbers — is what lets AI-based simulations test theories about dark matter by modeling how billions of simulated particles would behave under different assumptions, then comparing the result to what telescopes actually observe.
The lesson's black hole example deserves a teacher's honest footnote. The first-ever image of a black hole (2019, the M87 galaxy) was built by combining data from radio telescopes on different continents into one virtual Earth-sized telescope, then reconstructing an image from that combined data using computational algorithms — some of that reconstruction work is accurately described as AI or machine-learning-assisted (a follow-up 2023 reprocessing of the same data, called PRIMO, explicitly used machine learning trained on simulated black holes to sharpen the image). It is fair to teach this as "scientists used AI to help turn telescope data into a picture," but it wasn't a camera snapping a photo — it was months of computational reconstruction from radio-wave data invisible to human eyes.
Climate and earth science. Satellites constantly photograph the Earth's oceans, ice, and forests. AI models trained on labeled satellite imagery can scan that firehose of images far faster than any team of scientists and flag things like coral bleaching, rising ocean temperatures, or the first faint heat signature of a wildfire — often catching a wildfire within minutes of ignition. Weather prediction works similarly but on atmospheric data: models like DeepMind's GraphCast are trained on decades of historical weather patterns and learn to predict how weather tends to evolve, which lets them generate a 10-day forecast in seconds rather than the hours a traditional physics-based supercomputer simulation takes — and, on many measures, more accurately.
Biology and chemistry. A protein is a long chain of amino acids that folds into a specific 3D shape, and that shape determines what the protein does in the body — figuring out the shape from the chain used to take years of lab work per protein. DeepMind's AlphaFold was trained on the (relatively small) set of protein shapes that had already been painstakingly solved in labs, and learned to predict the shape of a new amino acid sequence directly — it has since produced predicted structures for essentially every protein known to science, over 200 million of them, which the field genuinely regards as a 50-year grand challenge being solved. The materials-discovery example works the same way: a model trained on known stable crystal structures can propose new combinations of elements likely to form stable materials, which researchers then have to actually synthesize and test in a lab — the AI's ~380,000 candidates are proposals, not finished materials.
One thing worth being upfront about, especially with older students: these are real, documented uses of AI in science, not speculative "AI of the future" content — but the specific numbers used as classroom hooks (200 million proteins, 380,000 materials, "10 days ahead") are illustrative of the scale these tools operate at, not live statistics to quiz students on precisely. Treat them as "here is how big this got," not exam trivia.
A second thing worth anticipating, especially with Hacker and Architect classes: a curious student may ask whether an AI that's this good at prediction actually "understands" the science. The honest answer is no — these models learn statistical relationships between inputs and outputs from historical data; they don't reason about *why* a star dims, a hurricane forms, or a protein folds the way it does. That gap between powerful prediction and genuine understanding is a real, ongoing conversation in the scientific community, and it's exactly what the 11–14 and 15–18 quizzes are pointing at when they ask about "over-reliance on correlations without understanding causal mechanisms."
Nothing beyond what AIQ normally needs: one device per student (or per small group) with a modern browser and an internet connection to open aiq.ph. No login is required to work through the lesson, no printouts, no lab equipment, and no advance setup — everything is self-contained inside the app. If you plan to run the discussion or wrap-up as a whole-class conversation rather than silent reading, budget an extra 5–10 minutes beyond the in-app content, and it helps to have the board or a shared screen free for jotting down student examples.
If your classroom has only one shared device or a projector rather than one device per student, the lesson works fine delivered as a group read-along — open it on the shared screen, read each scene aloud, and use the age-specific activity scripts linked below as your delivery guide rather than relying on students clicking through it themselves. This works especially well for Explorer and Builder mode, where teacher-led pacing is closer to how the lesson is meant to land anyway. For Hacker and Architect mode, individual device access is more valuable, since parts of the activity ask students to read and reason at their own pace before the group debrief.
No specialized vocabulary needs to be pre-taught. Terms like planet, telescope, weather, ocean, protein, and microscope are things students at every age band in this range have already encountered in a regular science class. The one piece of vocabulary worth previewing for Hacker and Architect classes, if you have a minute before starting, is "training data" — the set of past examples an AI model learns from — since both of those quizzes build directly on that idea, and it connects this lesson back to whatever AIQ lesson on machine learning fundamentals your class covered earlier in World 1 or 2.