Skip to main content

In-Class Studio: AI Visual Audit

Where this fits

You have the worked example (viz03), the audit categories (viz04), and the AI-prompt pattern (viz05). This page is the in-class studio where you put all three to work: read a deliberately flawed plot, identify at least three problems, revise the plot, and write a short audit note. The same workflow is what Assignment 5 expects, so think of this as the dress rehearsal.

Before you start

Take 2 minutes to check that your Codespace is ready:

  1. library(tidyverse) runs without error.
  2. The cached data file resolves from the repository root: examples/nhanes-equity/data/nhanes_equity_v6.csv.
  3. The rendered Week 6 worked example opens in the course book; you do not need its source file in your personal workspace.

If any of those three fail, raise it with your debugging partner before moving on.

Goal

Evaluate one AI-generated visualization for correctness, clarity, bias, aggregation choices, and misleading design. Then revise it so the visual matches the evidence.

Inputs

Activity Steps

  1. Review the worked example and note how it labels the data source, statistic, and limitation.
  2. In your personal workspace, open assignments/assignment05-visualization/flawed-plot-code.R and read the comments and output labels before you run it.
  3. Run the script from your personal workspace root with source("assignments/assignment05-visualization/flawed-plot-code.R").
  4. Look at the figure. It may appear in a plot viewer pane. If it does not, no problem: the script also saves flawed-plot.png in your current working folder (run getwd() if you are unsure where that is). Find the PNG in the Explorer, right-click it, and choose Open Preview.
  5. List at least three issues that could mislead a reader. Tag each one with one of the audit categories below.
  6. Revise the plot so the title, scale, grouping, and caption match the evidence. Keep the cohort restriction explicit.
  7. Write a short note explaining what you changed and why.

Audit Categories

  • Correctness: Does the plot show what the title and labels claim?
  • Design: Do scale, color, ordering, and geometry support fair comparison?
  • Interpretation: Does the caption avoid causal or population-level claims that the plot cannot support?
  • Bias and aggregation: Could grouping, missingness, or small counts hide important variation?
  • Reproducibility: Can another person rerun the code from the repository root using relative paths?

Required Outputs (studio draft)

These are the draft files you produce in class. The Assignment 5 deliverable list (viz07) is the same set, polished.

  • visual-audit-note.md — at least three issues with category, why it matters, and the corrected approach.
  • visualization-audit.qmd — a rerunnable Quarto file that creates the corrected plot and explains the revision.
  • corrected-plot.png — the exported corrected plot.
  • caption.md — a caption that states the statistic, data source, exclusions, and limitations.
  • ai-use-note.md — if you used AI, explain what it helped with and how you checked the result.
TipWhat a strong audit-note entry looks like

Each entry in visual-audit-note.md should give this level of evidence. This example is generic — it is not one of the planted issues in this studio’s plot:

  • Issue: the draft layered geom_smooth(method = "lm") across NHANES cycles.
  • Category: design.
  • Why it matters: cycles are categorical labels, so a regression line implies a continuous trend the data cannot support.
  • Correction: replaced the smoother with geom_line() + geom_point() connecting the cycle summaries.

Completion Check

You are done with the studio when a peer can rerun your corrected plot, identify the main comparison without reading your code, and see at least one explicit limitation in the caption.

Where this goes next

Polish the studio drafts into the Assignment 5 submission. viz07 walks through the seven A5 files one by one and connects the work to the M1 visual plan due the same week.