Check an AI Chart Description Against the Values Behind It
Compare chart descriptions with axes, units, source values, reversals, and limitations, then choose concise alt text and a useful nearby explanation or data table.
‘Steady growth’ can be a misleading description of a line that rises, falls and finishes only slightly above its starting point. The words sound convincing until someone checks the values. An AI description should therefore be reviewed beside the chart and its underlying data, rather than polished in isolation.
Verify what was measured, for whom, over which period and in what units. Check every claimed trend and exception. Then decide how to give readers the chart's essential information in text: sometimes a short alternative is enough; sometimes it needs a nearby description or table.
Establish the chart's job on the page
Write down the question the graphic answers. It might compare categories, show a time series or display a relationship. Describing grid lines and colors does little if readers need to know which category is largest or where a trend reverses.
Keep the title, source and surrounding sentence beside your draft. If nearby text already explains the key result, the alternative can identify the graphic and direct readers to that explanation. If the result appears only visually, supply it in text. The W3C complex-image tutorial describes this combination of a brief image identification and a fuller text equivalent when needed.
Stay within the evidence. A line ending higher than it began does not by itself demonstrate continuous growth, explain a cause or justify a forecast.
Make a small evidence grid
Read labels and source values before interpreting the visual shape. Separate the checks so a plausible headline cannot conceal a unit or population error.
| Check | Evidence to record | Question for the AI draft |
|---|---|---|
| Subject and population | Measure and group covered | Does the prose describe the same thing? |
| Horizontal axis | Dates, categories or variable | Is the order and period correct? |
| Vertical axis | Unit, scale and quantity | Does the wording preserve the unit? |
| Legend | Names of every series | Have groups been merged or exchanged? |
| Pattern | Beginning, end, peaks, lows and reversals | Is the claimed direction present throughout? |
| Limitations | Estimates, gaps, exclusions and notes | Has an important qualification disappeared? |
Use the underlying table for calculations where possible. The numerical-comparison guide covers denominator and percentage-point errors. Here, also ask whether a reader without the graphic can understand the relevant pattern and inspect significant values.
Check units before naming a change
An axis labeled ‘orders per 1,000 visits’ cannot support a statement that sales revenue doubled. A declining month contradicts ‘improved every month.’ Check axis titles, ticks, units and notes, including changes to denominators. Do not derive an apparently precise figure from pixels when no precise figure is supplied.
Missing information needs equal attention. A blank month is not zero, and a gap does not authorize an invented smooth continuation. Include a series that complicates the headline. If the source offers no table and the chart cannot support a precise number, request data or restrict the wording to a visibly defensible observation.
Repair a fictional description step by step
The following invented teaching data describe a line chart called ‘Workshop sign-ups by month.’ They represent no study, customer result or evaluation of an AI tool. The horizontal axis is January through April; the vertical axis counts sign-ups.
| Month | Sign-ups |
|---|---|
| January | 40 |
| February | 52 |
| March | 46 |
| April | 48 |
Suppose the AI draft reads: ‘Workshop participation grew steadily from 40 to 52 people over four months, proving that the new invitation strategy worked.’
That sentence makes four different errors. It turns sign-ups into participation, although the table says nothing about attendance. It uses the peak, 52, as the ending value, although April is 48. It ignores the fall from February to March. Finally, it invents a strategy and claims causal proof without evidence of either.
| Draft wording | Check against the table | Revision decision |
|---|---|---|
| Participation | Measure is sign-ups. | Keep the measure's actual name. |
| Grew steadily | Values are 40, 52, 46, 48. | Mention the peak and reversal. |
| From 40 to 52 | April ends at 48. | Correct the endpoint. |
| Proving the strategy worked | No strategy or comparison is supplied. | Remove the explanation of cause. |
A brief alternative could say: ‘Line chart of workshop sign-ups from January to April. February has the highest count, 52; April ends at 48. A table below gives all values.’
A longer explanation could state: ‘January recorded 40 sign-ups, February 52, March 46 and April 48. February was the peak. April ended eight sign-ups above January.’ The increase of eight follows directly from subtraction. Its cause remains unknown from these data.
The edit succeeds because it changes the claim to fit the evidence. A fluent restatement of the original error would still misinform readers.
Divide information between the alternative and nearby content
A simple chart with a few values may fit in a concise alternative. A complex graphic usually needs a more navigable arrangement than an enormous alt attribute. Stanford's chart guidance and Harvard's data-visualization guidance describe options including captions, body text and data tables.
Use a short alternative to identify the type, subject and principal supported takeaway. Where a table or full description follows, say so. Use visible prose for an important comparison, reversal or limitation. It also helps readers who see the chart but need interpretive context.
Provide a labeled data table or accessible source when exact figures or several series matter. Identify missing values explicitly. UW–Madison's complex-image guidance shows how a summary and table can complement a brief alternative. Avoid making readers encounter the same long passage in the alt text, caption and body; give each component a purpose. For broader image decisions, see reviewing AI-generated alt text.
After the evidence review, a grammar and style pass can simplify an awkward description. Recheck any generated edit against the grid before accepting it. A smoother sentence can still swap series or drop the word ‘estimated.’
Inspect the final chart in its final setting
Confirm that the approved text is attached to the correct figure, the promised description or table is present, and any detailed-data link works for the intended audience. A layout replacement can leave a correct description attached to an obsolete chart. Compare the final version again.
Read the page in order. Headings and links should lead to the explanation; the screen reader navigation guide covers the surrounding text structure. This editorial review cannot certify code, keyboard operation, color contrast or the experience of assistive-technology users. Those require separate technical checks.
For a presentation, check the speaker notes against the slides as well. The spoken account should use the same chart version, values and qualifications.
Save the source, chart version, final wording and outstanding questions. If reliable evidence is missing, leave detailed numerical approval unresolved rather than guessing.
Questions about chart descriptions
Can AI produce chart alt text without a reviewer?
It can suggest text, but someone must compare it with labels, values, context and data. UW–Madison warns that generated descriptions can be incomplete or inaccurate. Fluency does not establish that this comparison occurred.
Must every number be included in the alt attribute?
A simple figure may fit. For a complex chart, identify it briefly and provide the essential detail in nearby prose, a table or a clearly linked longer description.
What should I do when the chart and source table differ?
Pause. Ask the owner which version is authoritative, and correct the table or chart before approving the description. Choose according to the evidence, rather than the more appealing story.
