Most schools are still arguing about whether students should be allowed to touch AI. We think that argument is already over. The tools are in every pocket, woven into search engines, homework apps, and writing software your child already uses. Pretending otherwise doesn’t protect anyone — it just leaves students to figure it out alone, in secret, with no one teaching them the difference between using a tool well and laundering its output as their own work.
So our posture is simple, and it has two halves. First, we teach AI literacy — how to prompt, how to interrogate, how to catch the machine when it lies. Second, we assess in ways AI can’t fake. A student demonstrates understanding out loud, at a bench, defending a real lab notebook and identifying an unknown mineral in front of a person. There is no prompt that reads a streak plate or scratches a hardness point for you. When the assessment is honest, the studying becomes honest too, and AI turns back into what it should have been all along: a tutor that never gets tired, not a ghostwriter.
The course’s AI posture
We treat AI the way a good earth science teacher treats a rock hammer. It is genuinely useful and genuinely capable of doing damage, and the answer to both facts is the same: instruction, not prohibition. A student who has never been taught how AI fails — how it invents citations, misidentifies a mineral with total confidence, mis-states which way a fault moved, and tells you what you want to hear — is far more dangerous to their own learning than one who has been shown exactly where the tool breaks.
Our aim is a student who can sit down with an AI assistant and treat it like a sharp, fast, slightly unreliable study partner: useful for drilling the rock cycle, useful for re-explaining plate tectonics three ways, never trusted on a map reading or a dating calculation without checking the arithmetic, and never — not once — allowed to stand in for the thinking the student is supposed to be doing. The line we draw is not about the tool. It is about whose understanding ends up in the work.
Encouraged vs. off-limits
Here is the bright line, stated plainly. The left column is AI used to build your understanding. The right column is AI used to replace it. The difference is not subtle, and your child will learn to feel it.
| ✓ Encouraged | ✗ Off-limits |
|---|---|
| Drilling facts you must know cold — ask AI to quiz you on the rock cycle, the Mohs hardness scale, or the geologic time scale until you can recite them without it. | Submitting AI text as your own lab notebook. The notebook is a record of what you observed and measured. Borrowed words describing a mineral ID you didn’t perform are a falsified record. |
| Re-explaining hard concepts — have AI explain why plates subduct, or what radiometric dating really means, three different ways until one lands. | Having AI interpret your data for you. Pasting your seismogram arrival times in and copying out the epicenter distance, without doing — and understanding — the triangulation yourself. |
| Checking your own work after you’ve done it — identify a mineral yourself from its properties, then ask AI to verify and explain any mistake. | Copying a map interpretation or answer without verifying. Pasting an AI result into your work without checking the contour lines or the units actually make sense — they often don’t. |
| Summarizing your own notes — paste in your notes on plate tectonics and ask AI to summarize, then check whether the summary matches what you meant. | Outsourcing the reasoning. Asking AI for the conclusion of a stream-table or plate-boundary problem you were assigned to reason through yourself. |
| Debugging your reasoning — show AI your line of thinking on a relative-dating problem and ask where the logic breaks, then judge whether it’s right. | Disguising the source. Editing AI output just enough to hide where it came from, then presenting it as original thought. |
Notice the pattern. Everything on the left ends with you doing the understanding. Everything on the right ends with the machine doing it for you and you taking the credit. When you’re unsure which column you’re in, ask one question: if the AI vanished right now, could I still identify this mineral and explain the result? If yes, you’re studying. If no, you’re cheating — mostly cheating yourself.
What AI simply cannot do
There is a hard floor under this whole course, and it is the bench. AI cannot feel a mineral’s greasy or glassy luster, cannot watch a drop of dilute acid fizz on limestone, cannot judge a streak’s true color against a porcelain plate. It cannot do the in-person demonstrations. No model can stand at the bench and identify an unknown mineral by streak, hardness, and acid test, then defend each call to an examiner asking follow-up questions.
That is the quiet genius of the model: when the finish line is a live demonstration, AI stops being a shortcut and becomes a training partner, because the only way it helps you is by getting you genuinely ready to stand at the bench yourself. We walk through exactly how this works in AI-proof by design — the design principle that lets us welcome the tool instead of fearing it.
An assessment you can fake with AI was probably an assessment that wasn’t measuring much to begin with. The mineral-ID defense doesn’t beat AI by being harder — it beats AI by being real.
Curated prompt library
Here are concrete prompts students can copy and paste to turn an AI assistant into an honest earth science study partner. The trick is to make the AI ask you things rather than tell you things. Notice that every one of these ends with you doing the work.
Save the ones that work for you. Over a semester, a student who studies this way builds something no AI can hand them: the reflex of explaining and reasoning out loud, which is exactly the reflex every demonstration rewards.
Checking the machine
Here is the single most important habit we teach: AI is confidently wrong, and in earth science it is wrong in specific ways. It will misidentify a mineral with total confidence, invert a sequence of rock layers, drop a unit conversion in a dating calculation, or state a landform’s origin as fact when it’s guessing. It states all of this in the same calm, authoritative tone it uses for facts, and that tone is engineered to be persuasive. A student who trusts it blindly will absorb errors that sound right — and a wrong identification that goes unchecked quietly becomes the foundation for the next wrong conclusion.
So treat every AI claim as a hypothesis, not a verdict. When the machine gives you a number, do three things: redo the arithmetic yourself, check the reading against the actual specimen or map, and never take an identification on faith without running the test yourself.
- Check against the specimen. Before trusting any identification, run the streak, hardness, and acid tests yourself. AI gets mineral ID wrong often.
- Redo the units. Carry units through every calculation by hand — a confident answer in the wrong units is a wrong answer.
- Never take an identification on faith. If AI names a mineral or dates a layer, verify it against your own tests and the geologic map before you write it down. This is the one place a check is non-negotiable.
A student who leaves this course able to catch the machine has learned something more durable than any single unit of earth science: how to think clearly in a world full of fluent, fast, confident voices that are sometimes simply wrong. That’s AI literacy. And it’s why we teach the tool instead of banning it.