Skip to content
Idaho AI Classroom
All lessons
AI ImpactEthics and Privacy9-12.AII.EP.3

The Cost of a Billion Examples

11–1250 minutes90 minutes (extend seminar and writing)

Standard quoted exactly

Explore the societal, environmental, and ethical implications of large-scale data collection and processing and how this relates to AI applications.

Student-friendly learning targets

  • I can connect large-scale data collection to the AI applications that depend on it, including energy, water, consent, and labeling labor.
  • I can explain why data-center loads matter in an arid Western state that is recruiting those facilities.
  • I can distinguish data I agreed to give a school from data a company infers or scrapes without a meaningful yes.

Essential questions

  1. What does a billion training examples actually cost in energy, water, consent, and human labeling?
  2. Who benefits when Idaho recruits data centers, and who competes for the same electricity and water?
  3. If an application is useful, does that settle the ethics of how its data was collected and processed?

Objectives

  1. Map an everyday AI application backward to data collection, processing, storage, and inference.
  2. Analyze competing claims on electricity and water in an arid Western grid that also irrigates farms and cools towns.
  3. Evaluate consent: terms of service, inferred data, scraped public posts, and school records under FERPA.
  4. Describe labeling labor as real work, often low-paid and sometimes psychologically costly, not as a magic pipeline.
  5. Write a position that holds usefulness and cost in the same paragraph, without boosterism or panic.

Key vocabulary

Training data
The examples a model is shown during development so it can learn patterns; scale is often described in billions of tokens or images.
Data center
A facility packed with servers that store data and run computation, including model training and inference, using large amounts of electricity and often water for cooling.
Inference
The work a trained model does when it answers a new prompt; each query has an energy cost, usually smaller than training but repeated millions of times.
Consent
A meaningful agreement to collect and use data. A buried terms-of-service clause is legally common and ethically thin.
Labeling labor
Human work of tagging, rating, transcribing, or filtering examples, including toxic content, so models can learn and so outputs can be moderated.
Scraping
Automated collection of data from websites or platforms, often without a person-by-person request.
Externalities
Costs borne by people who did not choose the transaction: drought-stressed aquifers, grid strain, or underpaid raters overseas.

Teacher background

Large models exist because large data collection and processing exist. This lesson asks students to explore societal, environmental, and ethical implications of that scale and to tie them back to applications they already use. Idaho is actively recruiting data-center investment. That is an economic development fact, not a verdict. Data centers bring tax base, construction, and a small number of permanent jobs. They also draw electricity on a grid that serves farms, homes, and industry, and many designs use water for cooling in a region that already irrigates from the Snake River Plain aquifer and watches snowpack. Students should be able to state those tensions without turning class into a rally. Consent is the second strand: public posts, scraped books, school-issued accounts, and inferred profiles are not the same kind of yes. Labeling labor is the third: people, often contractors far from Idaho, rate toxic content and tag images so that applications look smooth. Use public figures of record (utility integrated resource plans, county hearing notices, company sustainability pages) and label them as claims, not gospel. Do not assign students to scrape the web or to upload classmates’ posts. FERPA still forbids treating student records as convenient training fuel. Offline fallback: printed utility excerpts, a water-budget worksheet, and a labeling-labor testimony excerpt from a reputable news investigation.

Materials and prep

Materials

  • Three-strand packet: Energy and water / Consent / Labeling labor, with short dated excerpts.
  • Idaho water-and-power worksheet: a simplified budget with households, irrigation, existing industry, and a hypothetical large computing load. Numbers are rounded and labeled as a teaching model, not an audit.
  • Consent ladder poster: explicit yes, account terms, inferred data, scraped public posts, no notice.
  • One excerpt describing content-moderation or data-labeling work (teacher-selected, not graphic).
  • Offline fallback: no live maps that require student logins; printed facility-siting hearing excerpts if a local proposal exists, otherwise a generic arid-West scenario clearly marked as hypothetical.

Before class

  • Check whether a data-center proposal is currently before a nearby county. If yes, use public hearing documents. If no, use a labeled hypothetical so you do not invent a local deal.
  • Sanitize the labeling excerpt for graphic violence; the point is labor, not shock.
  • Coordinate with a science colleague if you want snowpack or irrigation numbers that match a recent water year.
  • Block plan: extend into a written policy brief for a county commissioner: benefits, grid and water costs, consent conditions, labor standards.

Instructional sequence

Bill the query

5 min
  1. Students estimate, on a sticky note, what one chatbot answer costs in electricity. Collect wild guesses without mocking them.
  2. Tell them honest published estimates vary and that the bigger civic question is scale: millions of queries, plus training, plus cooling, plus the buildings Idaho is recruiting.
  3. Write three words on the board: energy/water, consent, labor. Today’s exploration has to touch all three.
  4. Read the standard. Emphasize explore and how this relates to AI applications, not only to warehouses on the horizon.

From application back to the warehouse

10 min
  1. Pick one application students know: a photo search, a chatbot, a translation app. Draw backward: device, network, inference, training, data collection, labeling.
  2. Define data center in Idaho terms: siting near transmission, tax incentives, construction jobs, modest permanent staffing, large continuous load.
  3. Show the water tension without theater: irrigation on the Snake River Plain, municipal use, drought years, evaporative cooling. Competing goods, not a cartoon villain.
  4. Introduce the consent ladder and labeling labor as ethical, not only environmental, costs of the same pipeline.
  5. Remind students that student records are not public training data. FERPA is a consent floor, not a suggestion.

Run the Idaho budget together

10 min
  1. Project the simplified power-and-water worksheet. Allocate a hypothetical large load. Ask what gets delayed: a housing development, a cold-storage plant, a farm pump, a school?
  2. Force a tradeoff sentence: If commissioners say yes, the benefit is… and the competing use is…
  3. Place one consent example on the ladder: a public Instagram photo used in a vision model. Where is the yes?
  4. Read two sentences of the labeling excerpt. Ask what the application’s smoothness hid.

Three-strand brief

12 min
  1. Individually, students write a 12–15 line brief on one application (maps, chat, image tools, recommendation feeds) covering environment, consent, and labor.
  2. Require one dated source from the packet and one explicit uncertainty (what the packet does not prove).
  3. No empty adjectives: do not write devastating or revolutionary. Write quantities, parties, and unknowns.
  4. If a student finishes early, add who captures the tax benefit and who lives next to the substation.

Real-world examples

  • County hearings in the Treasure Valley or Magic Valley on large computing loads, transmission upgrades, and water rights, using only public documents.
  • Idaho Power and other utilities planning for industrial electrification and large new loads in integrated resource materials.
  • A farmer on the Snake River Plain competing for water in a low-snowpack year while a cooling system evaporates water elsewhere in the basin.
  • Contract raters overseas labeling toxic content so a U.S. chatbot can refuse it quickly.
  • A school-issued account whose clickstream is processed under a vendor contract the student never read.

Hands-on activity

Consent ladder and labor station

8 min
  1. Half the class sorts five data examples onto the consent ladder and must justify each placement in one sentence.
  2. The other half lists what a fair labeling job would include (pay, trauma support, time limits, the right to refuse a queue) and what applications would get slower if that labor were treated as skilled work.
  3. Swap for three minutes if time allows; otherwise jigsaw report out.
  4. Collect the briefs. Star any brief that related costs back to a specific application rather than to data centers in the abstract.

Discussion questions

  1. If a data center pays taxes and uses legal water rights, is the ethical question settled?
  2. Should people whose public posts were scraped be able to opt out of training, even if they posted in public?
  3. Who should pay for grid upgrades: ratepayers, the company, or the state that recruited the facility?
  4. Is labeling toxic content a job we should automate, a job we should improve, or both?
  5. Does a useful translation app change how you judge the warehouse that makes it cheap?

Differentiation

Support

  • Provide a brief template with three labeled boxes: Environment, Consent, Labor, plus a sentence starter for uncertainty.
  • Allow a partner for the worksheet math; the writing remains individual.

Challenge

  • Compare training cost versus inference cost and argue which policy lever (siting, pricing, model size, query limits) actually moves the total.
  • Read a utility excerpt and identify one assumption a student would want cross-examined at a hearing.

Multilingual learners

  • Pre-teach aquifer, grid, consent, and contractor. Allow the brief’s first draft in the student’s strongest language with an English glossary of the three strands.
  • Note that labeling labor is often multilingual work done far from the product’s users.

IEP / 504

  • Provide a calculator and a simplified worksheet with fewer line items.
  • Accept a structured oral brief recorded on a school device that does not sync to a consumer model.

Assessment

Formative

  • Warm-up estimates and tradeoff sentences during the worksheet.
  • Consent-ladder placements with a justifying clause.

Summative

  • 12–15 line three-strand brief scored on application link, dated source, named uncertainty, and civic tone.
  • Block extension: 400-word commissioner brief with a recommended condition (water reporting, labor standard, or consent rule).

Success criteria

  • Addresses environment, consent, and labeling labor, not only one.
  • Connects costs to an AI application, not only to a building.
  • States a benefit and a cost without slogans.

Responsible use, ethics, and privacy

Responsible use

Students do not scrape websites, ping data centers, or attempt to measure a facility. Public documents and teacher packets only. No student should enter a personal account password or download a dataset of social posts. Offline printed excerpts are the core materials.

Ethics

Usefulness of AI applications does not erase environmental, consent, or labor costs; costs do not erase usefulness. Civic tone means stating who benefits, who pays, and what is unknown. Do not assign collective guilt to students for using a search bar. Do not sell data centers as destiny.

Privacy

School records, counseling notes, and classwork are not training data for public models. FERPA requires parent or eligible-student rights over education records and limits disclosure. Vendor contracts may still process metadata; students should know the difference between a district agreement and a consumer app’s terms. Do not collect classmates’ posting history as a project.

Reflection

  1. Which strand (energy/water, consent, labor) did you know least about this morning?
  2. What number in the worksheet would you want a journalist to verify before a county vote?
  3. How will this change, if at all, which applications you treat as cheap?

Homework

Read one public document (utility FAQ, county agenda item, or company water statement) provided on paper or a district site. Annotate one benefit claim and one missing number. Bring questions, not a rant.

Closing

Hold up the three words again. A billion examples are not free. Collect briefs. Next class asks what happens when those examples were taken without asking writers, artists, photographers, and people who posted in public.

Extensions and cross-curricular links

Go further

  • 90-minute block: seminar on siting conditions, then a commissioner brief.
  • Guest from a utility, county planning, tribal water staff, or farm bureau, with a recorded backup for rural bandwidth days.
  • Compare two companies’ public water claims and list what a journalist would still need to verify.
  • Trace one school-approved app’s privacy policy for training-use language; report clauses, not rumors.
Environmental science
Water budgets, evaporative cooling, snowpack, and competing beneficial uses.
Economics / government
Tax incentives, ratepayers, local control of siting, and who captures benefits.
Statistics
Orders of magnitude, uncertainty, and why a single query cost does not describe a billion-example system.
Geography
Transmission corridors, arid-West climate, and why companies look at Idaho.