Before you trust an AI-generated chart: five checks to make
An AI-generated chart can look ready for a presentation before anyone has checked how its numbers were calculated. The colours are tidy, the title sounds confident and the line moves in the expected direction. Those details do not tell you whether the comparison is fair.
AI can help plan a chart, prepare calculations and draft an explanation. You still need a way to compare the result with the underlying data. This guide gives you five checks and three copyable prompts to help you do that, whether the chart came from a spreadsheet assistant or another analysis tool.
Use an approved tool with data you are authorised to share. Keep the source table and the chart together, and work on copies when the tool can make changes.
Start with a question you can check
Consider this example: a product team wants to show the proportion of eligible onboarding requests completed in a month. One group completed 9 of 10 requests; another completed 10 of 100. An assistant has proposed an overall completion rate of 50%.
| Group | Completed | Eligible | Rate |
|---|---|---|---|
| Small group | 9 | 10 | 90% |
| Large group | 10 | 100 | 10% |
| Combined | 19 | 110 | 17.3% |
The proposed 50% is the simple average of 90% and 10%. It answers a different question: the average of the two group rates, giving each group equal weight. For the proportion of all eligible requests completed, the calculation is 19 ÷ 110 × 100 = 17.3%, rounded to one decimal place.
That difference would change the message in a dashboard or stakeholder update. Start by defining the question and calculation before choosing colours or writing the headline.
Five checks before you share an AI-generated chart
1. Trace the plotted values to the source
Purpose: Check what data the chart actually uses.
Find the source filename, sheet or table, selected range and reporting period. Confirm whether the chart includes every intended group and month. A filtered table or an old export can produce a plausible chart that answers the wrong question.
Compare plotted values with the source records. If the tool produced a summary table before plotting, check that intermediate table too. Ask for formulas or code so the calculation can be inspected separately from the image.
With only a screenshot, you may be able to check visible labels and axes. You cannot establish hidden filters, excluded rows or the underlying calculation. Record those checks as outstanding and request the source.
2. Recalculate the measure and denominator
Purpose: Make sure each value means what its label says.
The denominator is the total used as the base of a fraction or percentage. In the onboarding example, it is all 110 eligible requests. Dividing by completed requests or using only one group would describe something else.
Check whether the measure is a count, total, mean, median or percentage. A mean adds values and divides by their number; a median is the middle value, or the mean of the two middle values, when the observations are ordered. They can tell different stories when a few observations are unusually large.
For combined rates, add the relevant counts before dividing when they represent compatible, non-overlapping groups. Do not add group counts if the same request can appear in several groups. If group definitions or eligibility rules changed, resolve that before comparing periods.
Recalculate from the original counts rather than rounded labels. In the example, keep 19 and 110 in the working table, then round the displayed percentage at the end.
3. Inspect the chart type, axes and labels
Purpose: Check whether the visual represents the numbers fairly.
A bar chart can compare categories; a line chart can show a measure over ordered time periods. Check that dates run in the intended order and that gaps between observations are represented honestly.
The Office for National Statistics guidance on axes recommends a zero baseline for bar and area charts because their length or area represents the value. Line charts can sometimes use a shortened axis to make variation clearer; inspect the scale and its effect on the message. Keep scales consistent across panels intended for comparison.
Check units, legends and labels against the plotted values. A title saying “requests” above percentages is ambiguous. Make the chart readable on a phone, avoid colour as the only way to distinguish groups and provide an accompanying table or text description. The ONS guidance on chart text explains how titles, units, sources and descriptions help readers interpret the graphic.
4. Look for missing context and uneven comparisons
Purpose: Find what could change the interpretation.
Check missing months, incomplete periods and changes to the population being measured. A month-to-date count should not be presented as directly comparable with a completed month without a clear explanation.
For a percentage, show the underlying count where it matters. A 90% result based on 10 requests tells you something different about the amount of evidence available from a 90% result based on 1,000. It does not automatically establish a reliable difference between groups.
Keep blanks distinct from zeroes. If the values are estimates, preserve any supplied uncertainty information and explain relevant limits. Do not ask AI to invent confidence intervals or significance claims from a picture.
5. Check the headline and explanation
Purpose: Make sure the words stay within the evidence.
A chart can show that two things changed together without proving that one caused the other. If onboarding completion rose after a process change, other influences may also have changed. A suitable description states what the data shows and keeps the explanation as a question to investigate.
Check the difference between percentage points and percentage change. A rate moving from 20% to 25% rises by 5 percentage points. Relative to its starting value, it rises by 25%: (25 − 20) ÷ 20 × 100. Label the comparison you mean.
Remove language such as “significant improvement” unless the evidence supports the intended meaning. A short series of observations does not automatically establish a lasting trend. Ask what decision the chart can support and what remains uncertain.
Ask AI for an audit you can inspect
Use this prompt after you have a proposed chart and the source data. It asks for evidence for each plotted value, rather than a general assurance that the chart looks correct. Check the calculations yourself or with a separate spreadsheet formula, and have someone familiar with the measure review unresolved findings.
For the onboarding example, a useful audit would flag the 50% headline, show the 19 and 110 source counts and recommend a combined rate of 17.3%. It should also retain the group counts so readers can see why the overall result differs from the average of the two displayed rates.
Write the explanation after the checks
Once the chart is checked, ask for an accessible explanation linked to the verified figures. For the example, a supported takeaway is: “Across the two groups, 19 of 110 eligible onboarding requests were completed, giving a combined completion rate of 17.3%.” This describes the available data without inventing a reason for the result.
Try one check on your next chart
Choose a chart you are about to share. Trace one plotted value to its source, recalculate it and read the title again. Then use the remaining checks to inspect the full chart before it informs a decision.
Keep a short review note with the source version, calculation, filters, findings and reviewer. If a missing source or unresolved definition could change the message, leave the chart in draft until that point is resolved.
Sources checked 21 October 2026: ONS guidance on axes and gridlines and chart text, linked above. The onboarding example, arithmetic and prompts are teaching examples.
You May Also Like
How to contribute to a technical discussion when you do not know every term
How AI is reshaping development teams, from tickets to pull requests
