AI and Data: Spreadsheets & Visualization

Teacher note

Accounts to pre-stage

• The free Google account from earlier in the course covers Sheets today and NotebookLM next module. If somebody skipped it, get them set up in the first ten minutes, not at 0:20.
• One chatbot account each (made back in Module 1). Free tiers differ on file upload, so check that the tool a student is using actually accepts a CSV before the lab block.
• No new accounts are created today. That is deliberate, this lab is heavy enough.

Fifteen minutes before

• Copy the shelf files into the shared folder and open one to confirm it is not an HTML error page.
• Upload Mauna Loa annual CO₂ to your own chatbot once, so the first demo is not the first attempt.
• Open Guess the Correlation in a tab and play one round to clear any first-load prompts.
• Write the two column headings on the whiteboard: caught and missed.
• Put the wifi fallback USB stick on the desk where you can reach it.

Wifi fallback

Same files by USB, opened in local Sheets or Excel. The verification work is the point and it does not need the internet: the AI summary can come from a printed handout you generated earlier, and students check it against the rows in front of them. The lab survives, only the chart iteration gets slower.

Depends on earlier modules

• Module 1 gave everyone a chatbot account and the predict-the-next-word demo. Today is where that demo cashes out: an AI that recalls a number is guessing, an AI that computes one can show its work.
• Module 2 built the habit of iterating on a prompt instead of accepting the first answer.
• Module 3 ran the draw-a-scientist tally, so the room already knows how to argue about what a sample does and does not show.
• This session sets up Module 5. The verification habit gets pointed at citations next time.

Teacher note

The objectives above are the student-facing wording. The one you are actually teaching to is narrower: by the end of the night every student has stated one statistic out loud that they verified with their own hands. If that happened, the session worked, even if the charts are ugly.

Teacher note

Students see the same shape on their own page, so you can put either version on the projector. The lecture block is the one to guard: it wants to grow, and everything it takes comes out of the climate arc.

Teacher note

Only about seventy-five of the ninety-five minutes are structured. The slack is not a mistake, it is the wander block, and it is where a survey course gets personal. If the climate arc runs long, let it, and take the time out of the wander.

Anticipated wrong turns

• The summary describes a file that is not there. A model that cannot read the attachment will still answer, confidently, from the filename. This is the single best thing that can happen in the first fifteen minutes, so do not rescue it. Ask the room how we would have caught it.
• Nobody opens the rows. The summary sounds right, so the checking step gets skipped. Make it a called-on task, not an invitation.
• The free tier will not take an upload. Have a pasted-text version of the small file ready, 65 rows fits in a message.
• The file loads as one column. Tab versus comma separated. Fix it once on the projector, it will come up again in the next block.

Discussion, with the answers you are steering toward

"What did the AI catch that you would have missed?"

Usually the shape: row count, date range, a column with gaps. It is fast and genuinely useful at describing structure.

"What did it miss?"

Meaning. It does not know that a -99.99 is a missing-value code, or that two columns are the same measurement from different instruments. Context is the part you bring.

"It said the average is 355 ppm. How would you check that?"

Take the answer you want out loud: a formula in the sheet, or the AI's own method shown as a formula. Never a number with no method attached.

Anticipated wrong turns

• The file is too big to upload. The full per-country emissions file will not fit. Send them to the per-indicator file, or have them slim it in Sheets first. That filtering step is a real skill, so name it rather than treating it as a detour.
• A dramatic chart with a chopped axis. Ask what the chart looks like starting at zero, then ask which version is more honest. There is a defensible answer either way for anomaly data, and arguing it is the lesson.
• The AI does the arithmetic freehand. Trends and averages computed in prose are frequently wrong. Demand the method every time.
• Two sources disagree, so one must be lying. This is the stretch puzzle arriving early. Take it, it is the best moment in the block.
• Charting everything, verifying nothing. The most common failure. Circulate with one question: "which number here have you actually checked?"

Discussion, with the answers you are steering toward

"Your chart shows a rise. How much of that is the data and how much is the axis?"

Both, and that is fine as long as you can say which. An honest chart is one whose choices you can defend out loud.

"Two agencies, two numbers, same year. Who is wrong?"

Neither. Different baselines. Read the label before you read the number, every time.

"What would change your mind about this data?"

Push for something specific and checkable: a third independent source disagreeing, a break in the instrument record, a station that moved. Answers like "more research" do not count.

Anticipated wrong turns

• The county is not in the file. The civic shelf leans on California open data. If your room is elsewhere, swap in your own state's portal before class, and expect at least one student from out of state either way. Pair them with someone local for the block.
• Name mismatches. "St. Louis" versus "Saint Louis", counties with the same name in two states. A good five-minute lesson in why joining data is harder than it looks.
• Comparing dollars across decades. Nobody adjusts for inflation unless you ask. Ask.
• Over-reading a small county. A swing in a place with twelve thousand people is not the same finding as a swing in a place with two million.
• It gets personal. Unemployment and income data can be somebody's actual life. Keep the framing on the county, never on the person, and let anyone opt into a different county without explaining why.

Discussion, with the answers you are steering toward

"Did the number match what it feels like to live there?"

Often not, and that gap is the interesting part. Data about people is a summary of people, and summaries drop things.

"Whose decisions does this file feed?"

Funding formulas, business siting, benefit thresholds. The income-limits file is a good one to name here: it decides what counts as low income for a family of four in your county.

Spotlight sweeplab 1:25 – 1:35

Three or four students, thirty to sixty seconds each, chart on the projector. Scout them while you circulate, take volunteers first, then a gentle "that is great, will you show it?" tap. Never a cold call.

Prompts to hand the student showing

• "Say your one verified statistic, and say how you checked it."
• "What did the AI get wrong on the way here?"
• "Why this chart type and not a different one?"
• "What is the one thing this chart does not show?"

Prompts for the room

• "Who else picked this dataset and got a different answer?"
• "What would you want to see before you sent this to a boss?"

Wander block menulab 1:05 – 1:25

Twenty unstructured minutes with both instructors circulating. Offer the menu, do not assign from it. It also doubles as the shock absorber when an earlier block runs long.

• Start the budget builder. Mock transactions in, AI categorizes them, one total verified by hand. The highest buy-in thing in the module, because they will use it on Friday.
• Fact-check the narrative. For teams who want a race: a confident AI-written report meets the dataset it claims to describe, and every numeric claim gets logged as supported, unsupported, or distorted.
• A second dataset, chosen purely because it is interesting. Methane, sea level, life expectancy.
• The big-file exercise. Download something too large to upload, slim it to twenty rows in Sheets, then upload the subset. Real analysts do this daily.
• Spurious Correlations and the games shelf, for anyone who wants the warm-up game to keep going.
• The standing discussion: what would change your mind about this data?

Synthesis, five minutes

Go back to the two whiteboard lists from the first block, caught and missed, now that the room has three more hours of evidence. Add to both. The list of misses is the reason the checking habit exists, and it is much more persuasive in the students' own handwriting than in yours.

Exit ticket

Muddiest point, anonymous, one line on a slip or a form: what is still unclear about checking a number? Rotate the format across sessions so it does not turn into homework theatre. Read them before you plan Module 5, since the fabricated-citation lab lands better when it answers something the room already asked.

Homework handoff

The half-page assessment goes home: their own conclusion, in their own words, citing their own chart and their own verified number. Say the agency line once and move on: the data is public, the judgment is theirs.

Preview Module 5 in one sentence: next time the thing being checked is a citation, and the fabrication rate is one the class measures itself.