Your weather app is lying to you.
Not maliciously. It just has zero idea what to do when the weather does something it's never seen before.
A study published late last month in Science Advances tested the world's best AI weather models — Google DeepMind's GraphCast, Huawei's Pangu-Weather, and Fuxi — against the traditional physics-based model meteorologists have used for decades. The AI models crush it on average days. Faster, cheaper, often more accurate for routine forecasts.
But for the events that actually kill people — record-shattering heat, unprecedented cold snaps, wind speeds nobody's measured before — the physics model wins every single time.
The thing AI weather models can't do
AI models learn from historical data (1979–2017). They're insanely good at pattern-matching what's happened before. But when the weather does something that's never happened — which is happening more and more in a warming climate — the AI has nothing to pull from. It hits a ceiling. It can't imagine temperatures beyond what it's seen.
Physics doesn't have that limit. The equations governing atmospheric dynamics — Navier-Stokes, thermodynamic energy balance — don't care if a number has never appeared in a dataset. They extrapolate. That's the whole point.
The researchers found that AI models systematically underpredict the intensity of new records. Every time. Across heat, cold, and wind. Not by a little — by a lot.
Here's why your students need to hear: when a kid says "why do we need the equations? Can't AI just do it?" — this paper is your answer. AI interpolates. Physics extrapolates. And when the stakes are highest — disaster warnings, record-breaking storms — you need the model that can imagine what's never happened.
MORNING BELL RINGER
Ask: "How many of you checked a weather app today?"
Tell them: "A new study found AI weather apps miss record-breaking heat waves — the physics equations still win."
Ask: "Why would a physics model handle something it's never seen better than an AI trained on 40 years of data?"
Let them wrestle with it. That's the whole lesson.
Five minutes. Zero prep. Maximum "wait, WHAT?" energy.
FREE RESOURCE
Want your students arguing about AI vs. physics?
"Would You Rather: Trust AI or Physics?" — three forecasts, two methods, five-minute warm-up. Comes with a teacher answer key, a misconceptions list, and a sub-friendly read-aloud script (so the next time you're out, your sub can actually teach physics for once).
🚨 Quick Reminder
If Tuesday's Teacher Appreciation email hit your inbox during 4th period and you forgot — the 25% off code is still live through Sunday.
Everything in the store. No minimum. Applies to bundles AND merch.
Code: THANKATEACHER | Ends Sunday May 10 at 11:59 PM PT
Class Closer
The equations behind your weather forecast — Navier-Stokes — have never been fully solved. In 160 years. Mathematicians offer a million dollars to anyone who can prove they always work. Your weather app doesn't know this. It just says "bring a jacket.
P.S. What's your best "AI got it wrong" moment in class? We've all had that kid who breaks the simulation by entering a number the program didn't expect. Hit reply — we read every one.
Stay Wildly Curious,
— Lauren & Shawn
