Teaching the Water Question
The most transferable material in this section. Global versus watershed resources is a good science and civics unit, and it works whether or not the class ever mentions AI.
Why this is hard to teach.
This is the guide where “it's local” gets heard as “it's fine,” and it is the reason the whole section exists. The claim is accurate and it sounds like a defence. Read the three-beat rule linked at the bottom of this page before you teach this one — it is not optional here.
The second difficulty is that students find “local” and “small” hard to hold apart. Local means the answer depends on which basin you're in. In some basins the answer is serious. A facility that is a rounding error in a wet basin with spare treatment capacity is a real problem in southern Arizona, and both statements are the same claim.
Third, the honest finding of the main activity is an absence. Most students will look for their municipal system's large-customer withdrawal data and not find it. Students expect an activity to produce an answer, and this one produces a hole. Frame that as the result before they start, or half the room will think they did it wrong.
Target misconceptions.
“Local means small.”
Local means the answer depends on which basin you are standing in. Nationally the total is modest; in a stressed basin on a peak summer day it can be the binding constraint. Both are true, and that is what local means.
“Water used is water lost.”
Withdrawal is not consumption. Much of what is withdrawn returns to the basin. What leaves is what evaporates — roughly half, fleet-wide, though estimates vary and a widely quoted 80% figure is specific to cooling towers, not an industry average.
“The water is used at the data center.”
Most of it isn't. For one modeled GPT-3 output of a few hundred words: about 2.2 mL onsite and 14.7 mL at the power plant. Cutting a facility's onsite water to zero does not zero its water footprint.
“If it were a problem, someone would be measuring it.”
Operators do not publish facility-level water data, and in many states nothing requires them to. In North Carolina every large facility is on a municipal system, so no facility-level figure exists at all.
Two classroom-ready activities.
Facilitator warning
In drought-affected communities this material is not abstract, and the “it’s local” framing lands very differently in a room where the reservoir is visibly low. If that is your room, the three-beat rule is the difference between being useful and being resented. Lead with the stake, not the correction.
Trace your basin
40 min · public recordsStudents identify three things about where they live: which river basin or aquifer serves their county, whether it was under drought restrictions in any of the last three years, and whether their municipal water system publishes withdrawals by large industrial customer.
The first two are findable. The third usually is not — and that absence is the finding, not a failed activity. Say so before they start.
Debrief on the third one specifically. Ask what it would take to change it, and who would have to require it. This is the moment where an environmental lesson becomes a civics lesson, and it happens on its own if you let the silence sit.
Materials: internet access. State environmental agency and municipal utility sites only — no account required anywhere.
The retired statistic
30 min · media literacyGive students the claim they have almost certainly seen: that a conversation with an AI model consumes about a 500 ml bottle of water. Ask them to find, from the primary source, what it actually said.
What they will find: it described GPT-3, in 2023, at specific facilities, and it was modeled rather than metered. The same paper's replacement figure is about 16.9 mL for a 150–300 word output. The researchers did not retract it under pressure — the number was always narrower than its retelling.
This is a media-literacy exercise with a real, current, high-stakes example, and it transfers to every other claim students will meet this year. The question to end on is not “who lied?” — nobody did — but “where in the chain did the qualifiers fall off?”
Materials: the guide's sources block, which links the primary paper. Works as a homework assignment with a 15-minute debrief.
Discussion prompts.
Ordered from easy to charged.
- 1.Which basin do you live in? Had you ever needed to know before today?
- 2.What's the difference between water withdrawn and water consumed, and why does almost no coverage make it?
- 3.Why can the same facility be a serious problem in one place and irrelevant in another?
- 4.The 500 ml figure wasn't a lie — it was accurate about something narrower. How does a true statement become a false one?
- 5.Your municipal system may not track large-customer use. Who decided that, and who could change it?
- 6.Is “this is a local problem” a way of taking something seriously or a way of setting it aside? What makes the difference?
Seeing whether it landed.
The two-word test. Ask students to use “withdrawal” and “consumption” correctly in a sentence about something other than data centers — agriculture works well. If the distinction transfers, it stuck.
The basin brief. One page: their basin, its drought history, what is planned in it, and what they could not find out. The gaps section is the graded part.
The retelling trace. Give students any circulating statistic and ask them to find the primary source and list what qualifiers were dropped along the way.
When someone asks “is this proven?”
In teacher voice
"The water research is real and it's mostly modeled rather than metered — which means it's estimates built from published facility data, not meters someone read. That's a legitimate method and it's worth saying out loud. The number you've probably heard was retired by the people who produced it, and the replacement is much smaller. What we genuinely don't know is what any specific facility near you uses, because in most of the country nobody is required to produce that number. So when I say the honest answer is 'nobody knows,' I don't mean it's unknowable. I mean nobody has to tell us."
Delivering a correction to a room that cares
Every guide in this series involves telling people that a claim they hold is imprecise. The flagship educator page carries the three-beat rule — the sequence that keeps a correction about scale from being heard as indifference. Read it before you teach any of these.
The three-beat rule →Where this leads
Teaching this is a different skill than knowing it.
Teaching AI Well is the facilitator library behind CPAI's Certified Applied AI Trainer Program — ten lessons on how to teach AI honestly, including the material on this page.