AI reading a chart: the numbers on the axes can change the answer
A model may see a bar correctly and misread what it means. Check axes, units, legends, missing data and the limits of its conclusions.
Article language: English

In this article
From 80 to 100 enquiries. Suppose a studio received 80 messages one week and 100 the next. The difference is 20, or a quarter of the earlier number. Yet on a chart with a truncated axis, the second bar can look three times as long. An AI assistant describing the image should read the numbers before announcing an explosion of interest.
Evaluating this use takes more than asking whether a model "understands charts". Check several things separately: does it identify the series and unit, read the values correctly, and keep its conclusion within the data? It can transcribe two numbers accurately and still draw the wrong conclusion. It can also calculate a difference correctly for the wrong axis.
Ask for the basis of the answer first
Start with the question the analysis actually needs to answer. "Describe the chart" invites a broad commentary whose scope is difficult to assess. "By how much did enquiries change between the indicated weeks?" gives a specific reference point. Before the test, prepare the answer from the source data and identify which parts of the image support it.
Ask the model to name the series, the values it read, the unit and the period being examined. Short, verifiable statements are enough to compare with labels and data. If it answers about a different series, you can spot the error before reading a finished paragraph for a report.
For 80 and 100, expect those values and identification of the correct weeks. If the model cannot read a label, it should say so. "About 100" may be enough for a rough description, but not for a comparison requiring an exact difference. The criterion depends on the intended use, so agree it before the evaluation.
ChartQA research, published in 2022, includes questions requiring visual interpretation as well as logical and arithmetic operations. It helps distinguish tasks that a general request to "describe the chart" can easily treat as one ability.
If the chart rounds its numbers, decide how precise the answer needs to be. An image labelled "about 100" does not provide the same basis as a table with an exact value. Your own test set can include both. Assess whether the model preserves the precision of the material, rather than rewarding extra digits the image does not support.
Keep reading and calculation separate in the test report. A model might read 80 as 60 and then correctly calculate a change from 60. In that case, revising only the arithmetic instruction misses the cause. Check the inputs first, then the operation and the final sentence.
Bar length is no substitute for the scale
For our example, prepare two images with the same values. The first shows an axis from zero to 100. The second starts at 70. In that version, the visible segments from the lower bound to the values correspond to lengths of 10 and 30. The actual values are still 80 and 100. The presentation has changed, not the number of messages.
Ask both versions the same question. The difference in enquiries should remain the same. The model may note that the second presentation makes the change look more dramatic, but it should not triple the result on that basis. If the answer follows the visible segment rather than the scale, the test reveals a specific limit on the use case.
Also check whether equal distances on the axis represent equal differences in value. With another scale, a similar distance in the image can represent a different numerical relationship. Record the reading method and expected answer in the control set. Do not demand an exact calculation from the image alone if its markings cannot be recognised.
Misleading ChartQA treats reasoning about charts with misleading features as a separate evaluation problem. Do not transfer its results to your own model without a test. The narrower implication for test design is that an ordinary set of clear charts is not enough to assess resistance to changes in presentation.
A version with a reversed axis tests something else: does the model read the markings, or automatically interpret a line moving upward as an increase in value? State the direction of the numerical change in the expected answer. Describe the meaning for the process separately; a shorter response time and a larger number of enquiries do not represent the same improvement simply because both charts look favourable.
Do not assume every truncated axis reflects bad intent. It may help show small differences if the description makes the range clear. In an AI assessment, evaluate whether the answer accounts for those conditions. You can separately assess the chart's design and whether a reader understands the presentation.
Which series and which unit?
A separate test can include two series: one shows the number of enquiries, the other the time to first response in minutes. The lines may cross even though they represent different quantities and use separate axes. Their crossing in the image does not itself mean the values have become equal. Each line must be assigned to the right unit.
Ask for the series by name, not just by colour. Then swap the colours without changing the data and repeat the question. The expected numerical answer stays the same. If the model still describes the former "red line", check whether it read the current legend or relied on earlier context.
In another image, change the order of the legend entries. This is useful when the organisation publishes several layouts and the assistant will receive them without additional instruction. Do not treat every variation as an independent test of the whole analysis. Record that you are testing the connection between a name and its visual representation.

Check whether a series shows a result for each period or a cumulative total. One hundred enquiries collected up to the end of the second week does not mean 100 new enquiries arrived during that week. A label or data description should make this clear. If it does not, the model can present possible interpretations and ask for clarification instead of choosing one without a basis.
If you design charts aligned with a brand's visual language (in Polish), agree clear labels and ways to distinguish series. These also help when assessing AI answers, but do not automatically prove them correct. A beautifully chosen burgundy does not excuse the model from reading the word "minutes".
A missing point is not zero
On a control chart, show eight enquiries on Monday, ten on Tuesday and twelve on Thursday. There is no reading for Wednesday. If a line connects Tuesday and Thursday, saying "there were eleven on Wednesday" would be an estimate from the line, not a value shown in the data. The model should name that difference.
Test two separate versions: an actual zero for Wednesday and missing Wednesday data. Expect different answers to a question about that day. The same applies to an empty cell in the source table. When turning it into an image, make sure the drawing process has not already converted the missing value to zero before the chart reaches the model.
Account for completeness when comparing weeks too. One hundred messages over seven days and 80 over three can be described as two totals, but not as comparable full weeks. Mark the incomplete period in the test set and check that the answer preserves it. Without further agreement, do not invent a result for the missing days.
The time range also limits the conclusion. A chart from June to August is not enough to assess a whole year. Cutting off the first months can change the visible direction without changing any displayed point. Prepare a full and a shortened version, and record in the expected answer which period each comparison covers.
Read the order of labels too. Bars sorted from largest to smallest may represent days of the week, but their arrangement is not chronological. The model should describe values assigned to days, not infer a trend from left-to-right movement. Keep the data the same and change only the category ordering in the test.
A similar limit arises with long context and information omitted by AI (in Polish). Here the limit may be the image crop or visible period. If the model has not received the whole basis for comparison, it should restrict the answer to what it has rather than fill the rest with a convincing story.
A test set should separate causes of error
Prepare charts from data for which you have an independently established answer. Change one feature at a time: axis range, colours, legend, crop or presence of a point. If you change five things together, a difference in answers will not identify the cause. A test may reveal a problem without showing what needs correction.
Keep the image exactly as the tool received it. Numbers legible in the original file may become unreadable after resizing. If text recognition runs before AI, compare that output too. Checking OCR before AI analysis (in Polish) helps distinguish a character-reading error from a later interpretation error.
Do not put the correct answer in the filename or accompanying description if you are assessing image reading alone. If the intended workflow also gives the tool a table, you may test that broader version, but name it separately. A correct answer after receiving tabular data does not prove the same thing as a correct reading of the chart.
Separate data transcription, calculation and interpretation in the report. When an assistant returns a structured format, check what its fields mean too. Valid JSON does not guarantee valid content (in Polish). A field called "change" may contain a difference in the number of enquiries or a relative increase; its name should distinguish them.
If the comparison is to assess an experiment, there is another layer. A chart alone does not settle whether a difference was caused by the change. Checking an AI description of an A/B test (in Polish) explores that question. In this assessment, focus on whether the commentary is properly grounded in the data shown.
A specific sentence reaches the reader
When an image is illegible, test a refusal to give an exact reading too. A useful answer may identify the label or unit it cannot recognise and ask for a better version. That is more useful than a general "I cannot analyse charts", and more honest than guessing a number. Record a correctly identified limit separately from a confidently stated incorrect interpretation.
The final assessment should say which answers can be accepted and when manual checking or the source file is needed. A tool may read labelled values well but require checking with dual axes. Do not reduce such a result to a general "understands" or "does not understand charts".
Aeneia Audit can assess how AI is used, its risks and acceptance criteria for this analysis. The scope should include charts actually used by the organisation. Evaluating one answer does not replace examination of the methodology behind the data or verify all future model commentary.
Before publication, return to the sentence the reader will actually see. In our example it may describe an increase from 80 to 100 enquiries in the two specified weeks. If it says "interest tripled throughout the season", correct it. Readers need two values with the right dates. A whole season requires separate data.










