An AI-generated chart can look ready for a presentation while showing the wrong denominator, omitting a category, or exaggerating a difference. Before you share it, compare the visual with the underlying data and check whether its headline says only what those data support.
This practical review works whether an assistant created a chart image, wrote plotting code, or configured a spreadsheet chart. None of those methods guarantees accuracy. Keep the source table beside the final export and work through the checks below.
1. Establish what the chart is supposed to show
Write down the question before inspecting the design: “Which workshop format did respondents select?” is more precise than “Show our survey results.” Then identify the population, measurement, time period, and any filters.
Ask for the exact table used to plot the chart, including calculated columns. If you supplied a file, compare that table with your original rather than accepting another AI-generated table as independent evidence. Check which sheet, range, or records were selected. A correct chart of the wrong subset still answers the wrong question.
If the chart came from a generated illustration rather than data-linked plotting, do not assume the bar lengths correspond to the printed numbers. Rebuild quantitative visuals from verified values when you cannot inspect how they were drawn.
2. Work through a deliberately incorrect example
The following data and flawed chart are fictional teaching examples, not survey findings, a product test, or a real benchmark. Imagine 100 people each selected one preferred workshop format.
| Format | Responses | Share of all responses |
|---|---|---|
| In person | 40 | 40% |
| Live online | 30 | 30% |
| Recorded | 20 | 20% |
| No preference | 10 | 10% |
| Total | 100 | 100% |
The flawed chart
Suppose the AI describes a horizontal bar chart titled “Most people prefer in-person workshops.” It shows only In person, Live online, and Recorded, with values 40, 30, and 20. It labels these 50%, 30%, and 20%, says “Total: 90,” starts the response-count axis at 15, and adds “Source: latest survey.”
Reconcile the numbers before fixing the styling
- Total: 40 + 30 + 20 + 10 = 100. The displayed bars sum to 90 because one category is missing. That subtotal cannot silently replace the full sample.
- Denominator: In person represents 40 ÷ 100 × 100 = 40%, not 50%. If the chart intentionally excludes “No preference,” the shares among the remaining 90 responses become approximately 44.4%, 33.3%, and 22.2%. Explain that filter and denominator explicitly.
- Rounding: Those rounded subgroup percentages total 99.9%. A rounding note can explain this small discrepancy; it cannot explain the original incorrect 50% label.
- Claim: In person is the largest single category, but 40% is not a majority. A defensible title is “In person is the most-selected workshop format in this fictional example.”
The corrected chart includes all four categories, uses response counts consistently, starts its numeric axis at zero, and labels the total as 100. Its source note identifies the fictional table above rather than inventing an external survey.
3. Check units, scales, and visual proportions
Read every axis and data label aloud. Are the values counts, percentages, minutes, or thousands of units? A count of 40 does not become 40% unless the denominator supports it. If monetary amounts appear in another task, verify the currency code and scale; a bare dollar sign or an unexplained conversion is not enough.
For the fictional bar chart, an axis starting at 15 makes the visible lengths for 40 and 20 equal to 25 and 5. One bar therefore appears five times as long even though the underlying count is only twice as large.
The UK Office for National Statistics guidance on axes recommends zero baselines for bar and area charts. It distinguishes these from line charts and scatter plots, where a cropped axis can help reveal variation. Do not apply a blanket “every axis must start at zero” rule.
Check tick spacing, reversed axes, log scales, and whether related panels use comparable scales. Inspect the plotted marks as well as their labels: a correct number printed beside an incorrectly sized bar does not repair the visual.
4. Look for missing categories and outdated sources
Compare the category list with the original table, including “Other,” “Unknown,” zero values, and missing observations. Missing data are not automatically zero. Ask whether blank cells were dropped, duplicate records were counted, or categories were combined without explanation.
For time series, check that dates are in order and intervals are represented honestly. Verify whether partial periods are being compared with complete ones. Preserve uncertainty information when the source includes it; a cleaner-looking chart should not quietly remove an important qualification.
Then inspect the source itself. Record its actual title, link or file version, coverage period, and relevant update date. A newly generated chart may still contain old data. “Latest” needs verification against the source, not the assistant’s confidence. Our guide to verifying AI citations explains how to check the supporting references.
5. Review the headline and accessibility together
Read the headline, caption, and accompanying paragraph separately from the graphic. Do they change “largest category” into “majority,” imply causation from a comparison, or omit a limitation? Use the same discipline described in checking an AI summary for meaning changes.
Provide readable labels and do not make color the only way to identify a series. For a complex chart image, W3C’s guidance on complex images describes a short text alternative plus a fuller description of the essential information. A nearby accessible table and a concise explanation can help convey values and relationships.
For our corrected example, suitable alt text is: “Fictional workshop preferences: in person 40, live online 30, recorded 20, and no preference 10 responses; full data in the accompanying table.” Check that the exported labels remain legible at the size readers will actually see.
A final check before sharing
Recalculate the key numbers independently, match every category and unit to the source, inspect the actual export, and read the main claim once more. Asking AI to audit its own chart can help find issues, but it does not replace these checks. If you cannot reconcile the visual with its data, share the verified table instead and hold back the chart.