I asked ChatGPT to make a chart

OK, so I asked ChatGPT to become an instant data visualization expert and create a pie chart with nine slices, preferably to be implemented in Excel.

What it said

Initial assessment: the suitability of the pie chart

It started by deciding it needed to evaluate whether a pie chart even made sense:

Pie charts are popular but ineffective when there are many categories (nine, in this case) or small differences between values. Since each slice depends on an angle, human perception struggles to compare areas and angles accurately.

➡️ Conclusion: An expert would recognize that a pie chart with nine categories offers poor readability.

So, after explaining to me as if I were five years old that pie charts should not be used, it suggested three more effective alternatives: a horizontal bar chart, a 100% stacked bar chart, and a waffle/pictogram chart.

It then outlined the extreme conditions under which an expert might accept the shame of creating a pie chart.

If I insisted on a pie chart

An expert would only accept a pie chart if:

• The audience didn't need a precise comparison of values;
• The goal was to show a general sense of the data (e.g., "housing represents about a third of expenses").

In that case, it would apply design best practices:

1. Sort slices by size, clockwise: Start at the 12 o'clock position (top) with the most important category (Housing).
2. Avoid similar colors; use a perceptually uniform palette (e.g., Viridis or Tableau 10).
3. Avoid separate legends: use direct labels next to the slices with name and percentage.
4. Avoid exploded slices, which distort perception.
5. Add the total to the title: "Monthly Expense (Total: €1,800)".

An expert's recommended version

A horizontal bar chart sorted by value, with direct labels and neutral colors.

Example configuration (in Excel):

• Type: Horizontal bars.
• Sort descending (Housing → Others).
• X-axis with the € scale (0 to 700).
• Show labels to the right of each bar (e.g., "€650 – 36%").
• Soft colors, highlighting the main category.

Result

The chart ChatGPT generated is essentially something that looks like it came out of Excel 2003.

Assessment

The task was simple. In some "mechanical" parts, ChatGPT got it right. But fundamentally, it didn't behave like an expert—just like someone who had read a few poorly digested articles. Let’s look at some hits and misses.

Pie charts and bar charts are not equivalent

In my course, I emphasize the idea that every type of chart, and every way of designing it, conveys a different message. This advice to replace pie charts with bar charts is flawed, though it was highly popularized by Edward Tufte and, especially, by Stephen Few.

Another suggested alternative is a stacked bar chart. In reality, if we want to ensure reading precision, a stacked bar chart is less precise than a pie chart. There are studies that show this, but you only need to try identifying the 25%, 50%, and 75% marks to realize it’s easier in a pie chart.

It is the question we ask that drives the design.

Insisting on the pie chart

This part is quite funny. ChatGPT basically says, "You’re an idiot, but if you really want to make a pie chart, here is how you should do it." The recommendations are exactly what anyone who has read one or two data viz books expects to find. But again, they are poorly digested:

• Sort segments by value: Of course, that’s the bare minimum.
• Avoid similar colors and use a perceptually uniform palette: Actually, in a pie chart, segments can even share the same color since slices are identified directly without a legend. Suggesting the Viridis palette is silly (segments are categorical, not continuous).
• Avoid legends and exploded segments: OK, data visualization 1.0.1.
• Add the total to the title: Probably not; if it’s important, there are better solutions.

One of the best reasons to use a pie chart escapes ChatGPT: showing the cumulative proportions of two or more slices.

An expert's recommended version

According to ChatGPT, an expert's recommendation would be the bar chart shown above. In truth, the way it described it would create a better-designed chart than the one it presented. But who hasn't been there, having a description better than the implementation?

Note that in the chart, the "Others" category tends to get the opposite treatment of the relevant categories: painted grey and placed last.

An interesting point ChatGPT makes is to emphasize the most important category (Housing) using a different color or tone and adding the annotation "Housing represents more than a third of monthly expenses." Annotations should refer to something visible in the chart. However, because it decided the bar chart was the most suitable, the conclusion that "Housing represents more than a third of monthly expenses" cannot be drawn from the chart because a bar chart does not allow for that conclusion. Capisce, ChatGPT?

In reality, a major part of the process is missing: the largest value isn't necessarily the most important. The most important is the one we want to emphasize (it could be, for example, the proportion of transport in a household budget). Designing the message also requires decisions about the nine categories: are they all relevant, or can we aggregate some of them? This is also not mentioned by our "expert" ChatGPT.

Conclusion

In a simple example, ChatGPT failed on the essentials: a chart is defined by the message, not by substituting one chart for another that doesn't say the same thing. It also failed by not proposing ways to read and interpret the data, including determining if all categories were necessary. It failed to mention one of the main advantages of pie charts over bar charts (evaluating the cumulative percentage of two or more slices). Finally, it claimed the bar chart was the right choice but then suggested an annotation that only makes sense in a pie chart.

For such a tiny problem, ChatGPT was far from being an expert. What will it do if we ask it for something more complex?

PS: I asked ChatGPT to evaluate this text, and it found it brilliant: "High-quality text, with technical authority, controlled irony, and strong pedagogical value." Which goes to show it doesn't always get it wrong...