# How do you verify an AI-generated chart or statistic?

Date: 2025-06-21 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/verify-ai-generated-chart-statistic/

**Answer:** Trace every number back to where it came from: the rows of your own file or a named, checkable source. Then recompute one or two key figures yourself, read the code the AI ran, and look at the chart's axes, because a correct number can still be drawn in a misleading way.

AI assistants have become very good at turning a spreadsheet into a clean chart with a confident headline. ChatGPT, for example, analyzes uploaded data by writing and running Python code, and it can merge files, clean them and draw bar, line, pie or scatter charts in the conversation. That is genuinely useful. It also means a polished picture can arrive before anyone has checked the numbers underneath.

There are two separate questions to ask of any AI-made chart or statistic. Is the number right? And is the picture fair? A figure can be computed correctly and still be drawn to exaggerate, and a beautiful chart can rest on a row that was silently dropped. *A Beginner's Guide to ChatGPT* gives the rule that covers both: models can make errors, so data and numbers always get a human check.

## Two kinds of AI numbers

Before checking, sort the number into one of two types.

**Computed from your data.** You uploaded a file and the AI calculated a total, an average or a percentage change. This is the easier case: the raw data is in front of you, and the AI usually ran code you can read.

**Recalled or found.** You asked a general question ("What share of small businesses use AI?") and got a figure. Here the number either came from a search result, which should be cited, or from the model's memory of its training text, which can be outdated or simply wrong. A statistic with no named source cannot be verified at all, so treat it as a lead to look up, not a fact.

## A worked check, step by step

A volunteer coordinator uploads a small spreadsheet of monthly visitors to a community library and asks for a chart for the newsletter. The data and the AI's output here are illustrative.

| Month | Visitors |
|---|---|
| January | 1,210 |
| February | 1,180 |
| March | 1,260 |
| April | 1,240 |
| May | 1,330 |
| June | 1,390 |

The AI returns a line chart titled "Library visits soar 40% in six months," with a steep upward line.

**Step 1: recompute the headline.** From January (1,210) to June (1,390) the rise is 180 visitors. Divided by 1,210, that is about 14.9%, not 40%. A quick calculation in a spreadsheet or on a phone takes under a minute and has already caught the main error.

**Step 2: ask the AI to show its work.** Ask in plain words:

```prompt
Show me the Python code you used for the chart and the percentage, and list
the exact rows and values it used. Then recalculate the change from January
to June step by step, and tell me if it differs from your headline.
```

Read the code even if you are not a programmer. Look for the rows it filtered, the columns it used and the formula. In cases like this one, the cause is often a wrong column, a duplicated row, or a comparison against the lowest month instead of the first.

**Step 3: check the axes.** The steep line comes from a y-axis that starts at 1,150 rather than zero. This is called a **truncated axis**, and it makes modest changes look dramatic. Research on such charts found that readers overestimate differences even when they are told the axis does not start at zero. Spreadsheet programs often truncate automatically when values sit in a narrow range, so this can happen without anyone intending it. Ask for the chart again with the axis starting at zero, or label the break clearly.

**Step 4: fix the words.** "Soar" is a loaded word, and loaded titles prime readers before they look at the data. A fair title might be: "Library visits up about 15% from January to June."

> **Try it:** Upload a small table you know well, such as household bills for six months, and ask an AI assistant for a chart with a headline figure. Recompute the headline yourself, then ask it to show its code and the rows it used. Note whether the axis starts at zero.

## A short checklist

| Check | Question to ask |
|---|---|
| Source | Can I point to the row, file or named publication behind this number? |
| Recompute | Do I get the same figure from the raw data? |
| Code | Which rows and columns did it use, and what did it filter out? |
| Axes | Does the y-axis start at zero, and is any break or log scale labeled? |
| Wording | Does the title describe the data, or sell it? |

## What should make you suspicious?

Round, dramatic numbers deserve a second look, especially when they come with words like "soar" or "collapse." So does any statistic offered without a source, or with a source you cannot open. If you ask where a figure came from and the answer changes, or a citation leads to a page that does not contain the number, set the figure aside.

Watch the shape of the chart, too. An axis that starts well above zero, a y-axis maximum chosen to flatten or steepen a line, missing tick marks, or an unlabeled logarithmic scale can all change the story without changing a single value. A logarithmic scale is legitimate for data that spans very different sizes, but unlabeled, it can make exponential growth look like a gentle straight line.

Finally, be wary of results that match exactly what you hoped to see. That is when people check least. The routine is the same every time: trace the number, recompute one figure, read the code, and look at the axes before the chart leaves your hands.

## Sources

- Improvements to data analysis in ChatGPT: https://openai.com/index/improvements-to-data-analysis-in-chatgpt/
- Misleading graph (revision of June 8, 2025): https://en.wikipedia.org/w/index.php?title=Misleading_graph&oldid=1294496743
- A Beginner's Guide to ChatGPT: https://www.amazon.com/dp/B0BZRRPXZW