How to Check an AI Chart Caption Against the Data, Line by Line
When an AI-written chart caption disagrees with its data, verify each claim against the chart’s source table and labels before editing the prose. Check the series, measure and units, dates, axis scale, totals, and uncertainty in that order. Record the caption’s exact claim, the evidence you checked, the mismatch, and the correction. This makes the review traceable and helps prevent a plausible-sounding summary from surviving a basic data error.
Start with the right version of the evidence
Collect the caption, the chart image or interactive chart, and the underlying data table or download. Also find the chart’s source note, subtitle, footnotes, and any methodology or definitions the source provides. Confirm that these materials refer to the same version, geography, time period, and measure. A chart image alone may not reveal the precise values; if the table is unavailable, mark claims that require exact values as unverified instead of estimating them from the picture.
Keep the AI’s draft intact while checking it. Break it into individual statements: which group is higher, when a change happened, how large it was, whether a total increased, or whether uncertainty ranges overlap. A single sentence can contain several claims, and each needs its own evidence. Official chart guidance treats titles, subtitles, sources, footnotes, axis text, dates, and category labels as meaningful parts of interpreting a chart, not decoration. Office for National Statistics (ONS), “Chart details: Chart text”
Match each series name to the plotted data
Compare every group named in the caption with the chart legend, direct labels, and corresponding column or field in the data. Check that labels have not been swapped, shortened into ambiguity, or confused with a related category. If the chart uses colors or line styles, trace each one to its legend entry; do not infer group identity from color alone when the mapping is unclear.
Then check the direction of the claim. If the caption says one series leads another, compare the values at the stated dates or categories. If it says a series rose, fell, or stayed level, check the relevant sequence rather than relying on the overall visual impression. Record the field name or label you matched so another person can repeat the check. ONS advises keeping category labels clear and concise, while retaining extra detail in surrounding text when labels are shortened. ONS chart text guidance
Verify the measure and units before comparing numbers
Read the axis title, tick labels, subtitle, and notes to identify what the values measure and in what units. A count, rate, percentage, and currency amount describe different quantities even when their numbers look similar. Check multipliers such as “in thousands,” percentage signs, and currency symbols. Make sure the caption does not turn a percentage-point change into a percent change, or describe a rate as a raw count.
When possible, compare like with like: the same measure, unit, population or category, and any stated denominator. If the caption uses a converted unit or calculates a difference, reproduce the conversion or subtraction from the displayed source values and note whether rounding explains a small discrepancy. Do not silently replace an unclear measure with a guessed interpretation. The U.S. Census Bureau’s reporting standard calls for labeled axes, units appropriate to the graph’s scale, and legends that define values or special terms. U.S. Census Bureau, “Statistical Quality Standard E2: Reporting Results”
Check dates, periods, and gaps
Compare every date named in the caption with the chart axis and source table. Distinguish a single observation date from a period such as a month, quarter, or year; check whether a year-ending period is labeled by its start or end year; and verify that the first and last points cited actually appear in the data. If the caption describes a change “since” a date, confirm both endpoints.
Look for missing observations, irregular intervals, or breaks in the series before describing a continuous trend. A line drawn between points does not establish that data were collected at every intermediate date. Check notes for unavailable or non-comparable estimates, and preserve such limits in the corrected caption. ONS guidance says date labels should identify dates or spans appropriately, use consistent abbreviations, and write dates out fully in chart metadata. ONS chart text guidance
Inspect the scale before describing the size of a change
Read the vertical-axis tick labels and determine whether the chart uses a linear or logarithmic scale and whether the displayed range starts at zero. Then ask whether the caption describes the data values or merely the visual size of a bar or slope. A truncated axis can make a small difference look large; a logarithmic scale changes how equal visual distances correspond to values. A caption should not call a change “dramatic” based only on the apparent steepness or bar height.
If multiple charts are compared, check whether their scales are consistent. If a chart uses a restricted range or a non-linear scale, describe the underlying values and avoid wording that exaggerates the visual effect. The Census Bureau recommends consistent scales for graphics likely to be compared and notes zero baselines as a way to avoid an inappropriate visual impression. Datawrapper’s line-chart guidance explains how linear and logarithmic scales map values differently and how custom axis ranges can be set. Census Bureau reporting standard, Datawrapper Academy, “Customizing your line chart”
Recalculate totals and derived claims
When a caption mentions a total, average, ranking, difference, or percentage change, trace it to the actual source values and calculation. Check which rows or periods are included, whether categories overlap, and whether the chart shows a subtotal or an overall total. Recalculate using the source’s stated definition and precision. If the chart provides rounded values, an apparent mismatch may result from rounding; do not claim an exact result that the displayed precision cannot support.
For a percentage change, identify the starting value and the comparison value, then check the denominator used in the calculation. For a total across categories, confirm that the categories are intended to be added and that no group is counted twice. If the source does not define a derived measure clearly enough to reproduce it, label the claim unresolved and seek the underlying documentation rather than supplying a plausible calculation.
Treat uncertainty as part of the claim
If the chart shows error bars, confidence intervals, ranges, or another uncertainty display, check the legend and notes to learn what the marks mean. Do not call an interval a confidence interval, margin of error, or standard deviation unless the source identifies it that way. Check whether the caption reports an estimate as exact, claims that two groups differ, or ignores uncertainty shown in the figure. When the source does not explain the interval, describe it neutrally as an uncertainty range only if that interpretation is supported by the chart materials; otherwise flag its meaning as unclear.
Uncertainty has a defined relationship to how a result is estimated and to the uncertainties of its inputs; the label should not be guessed from appearance alone. NIST’s summary distinguishes uncertainty components by how they are evaluated and explains that estimates derive from input estimates. For a caption review, the practical rule is to preserve the source’s terminology and avoid making a stronger claim than the estimate and its displayed uncertainty support. NIST, “Basic definitions of uncertainty”
Keep a discrepancy record and rewrite only verified claims
Use one record per claim so corrections remain auditable. A concise record can include: caption wording; chart or table location; series or category; measure and unit; date or period; source value or evidence; discrepancy type; resolution; and reviewer status. Useful discrepancy types include label mismatch, unit mismatch, wrong period, scale-driven overstatement, arithmetic error, uncertainty omitted, and not verifiable from available evidence. These are review fields, not a substitute for the source itself.
Once checked, rewrite the caption using only claims the evidence supports. Name the measure and period where they matter, describe the main observed pattern, and state important qualifications plainly. If a claim cannot be verified, remove it or say what remains unknown; do not preserve it because it sounds coherent. For readers who need the detail behind a chart, W3C recommends a short description paired with a longer description of essential information, which can include values, trends, and structured information such as a table. W3C Web Accessibility Initiative, “Complex Images”
A useful final check is to read each sentence alongside the matching source evidence and ask: could another reviewer locate the same value, label, date, scale, or uncertainty note? If not, the record or caption needs more precision. The goal is a caption that communicates what the chart supports and leaves unsupported interpretation out.