03 · What You Need to Know
Verify Both the Data and the Representation
A research table or figure has at least two layers of accuracy. The first concerns the underlying values. The second concerns how those values are represented.
Data accuracy
The values, statistics, categories, and calculations correspond to the underlying data and analysis.
Representational accuracy
The labels, scales, visual encodings, ordering, annotations, and other presentation choices communicate those values without materially changing their meaning.
A table can fail the first layer by reporting the wrong percentage. A graph can pass the first layer yet fail the second by plotting correct values against a misleading axis. Both require verification.
Current ICMJE recommendations also place responsibility for AI-assisted material on human authors and specifically state that AI use for data analysis or figure generation should be described in the Methods where applicable. Authors remain responsible for the accuracy and integrity of submitted material.
Start with the source data, not the finished display
The strongest routine check begins with whatever generated the table or figure: the cleaned dataset, statistical output, analysis script, spreadsheet, or other authoritative research record.
Do not verify a generated figure by comparing it only with a generated table if both were produced from the same unverified AI output. That can reproduce the same error in two formats.
Identify the data source and establish which variables, observations, groups, and analytical results should appear.
Verify which observations are represented
A table may look numerically consistent while being based on the wrong subset of data.
Check sample sizes, inclusion and exclusion criteria, group membership, missing-data handling, duplicate records, filters, and any other conditions that determine which observations entered the display.
If the table says n = 248, confirm where 248 came from. If a figure contains four groups, confirm that those groups correspond to the intended categories and that observations were assigned correctly.
Verify values against the analytical output
For each consequential value, identify its source.
Depending on the table or figure, this might include:
- frequencies and percentages;
- means, medians, and other descriptive statistics;
- standard deviations or other measures of variability;
- effect estimates;
- confidence or credible intervals;
- p-values;
- regression coefficients;
- model predictions;
- category counts;
- derived rates or ratios.
You do not necessarily need to manually reconstruct every cell in a large table. Verification effort should concentrate on consequential values while also using systematic checks to ensure that the complete output was generated from the correct analysis.
Recalculate derived values
Percentages, totals, rates, ratios, changes, normalized values, confidence intervals, and other derived quantities deserve independent checking.
If AI calculates that 72 of 120 participants represent 65%, the table contains a straightforward error even if everything around it is beautifully formatted.
For consequential numerical outputs, use the dedicated process for verifying calculations produced by AI. Check the inputs and formula rather than merely recalculating from values that the AI itself supplied.
Check denominators carefully
Percentages are especially vulnerable to denominator errors.
A percentage might use the full sample, valid responses only, one subgroup, participants completing follow-up, or some other denominator. AI may infer the wrong denominator when it is not explicitly provided.
The displayed percentage can even be mathematically correct while communicating the wrong quantity because the denominator is inappropriate.
Check totals and internal consistency
Tables provide several opportunities for simple integrity checks.
Do category counts sum to the stated total? Do percentages sum approximately to 100% when they should? Are subtotals consistent with group totals? Do confidence-interval limits surround their corresponding point estimates where expected? Does the number of observations match the analytical output?
Small differences can arise legitimately from rounding or overlapping categories, so discrepancies should be investigated rather than mechanically "corrected."
Verify row and column labels
A table can contain the correct values under the wrong headings.
Check variable names, categories, treatment groups, time points, units, reference groups, statistical measures, and any abbreviations. Pay particular attention when AI has rewritten technical labels to make them "clearer." Simplification can accidentally change meaning.
Check units
A value of 5.4 means very little without knowing what it measures.
Verify whether quantities are expressed in seconds or minutes, grams or kilograms, proportions or percentages, raw scores or standardized scores, thousands or individual units, and so forth.
If the underlying analysis transforms a variable, make sure the table or figure reflects the transformed scale correctly.
Verify axes and scales in figures
For graphs, inspect both axes carefully.
Check the minimum and maximum values, intervals between tick marks, scale type, units, category order, and whether the axis has been transformed. A logarithmic axis, for example, communicates distance differently from a linear axis and should be clearly represented.
A truncated axis is not automatically improper, but it can visually amplify small differences. The relevant question is whether the chosen scale helps readers interpret the data accurately rather than creating a stronger visual impression than the underlying effect warrants.
Verify that visual size corresponds to numerical magnitude
Bars, points, areas, bubbles, line positions, and other visual marks encode numerical information. Check that the encoding corresponds to the underlying values.
If a bar representing 60 is visually twice as long as one representing 50, the graphic can be misleading even if both labels display the correct numbers.
Three-dimensional effects, decorative perspective, and inconsistent scaling can similarly alter perceived magnitude without changing the printed values.
Check error bars and uncertainty indicators
If a figure contains error bars, bands, intervals, or other uncertainty indicators, establish exactly what they represent.
Are they standard deviations, standard errors, confidence intervals, credible intervals, prediction intervals, or something else? Were they calculated correctly? Do the legend and caption identify them accurately?
Nature's current formatting guidance, for example, requires error bars and statistics used in figures to be defined in the figure legend. This is publisher-specific guidance rather than a universal rule, but it illustrates why ambiguity about uncertainty indicators can undermine interpretation.
Check legends, annotations, and symbols
Verify every legend entry, color category, line type, symbol, significance marker, annotation, panel label, and abbreviation.
A legend that swaps treatment and control colors can reverse the apparent finding while leaving the plotted values untouched.
Nature's figure guidance similarly emphasizes clear axis labels, units, legible text, and accessible visual presentation. Again, individual journal specifications vary, so verify the requirements of the journal to which you plan to submit.
Check whether the table or figure shows what its title claims
A figure titled "Effect of Intervention X on Achievement" implies something stronger than "Mean Achievement Scores by Intervention Group" when the study does not support causal inference.
Titles and captions are part of the interpretation. Verify that they accurately describe the data and do not introduce causal, generalizing, or evaluative claims beyond what the analysis establishes.
Verify correspondence between the table, figure, text, and analysis
A common research-integrity problem is not necessarily that one element is wrong, but that different parts of the manuscript disagree.
Check that the values discussed in the Results match the table. Confirm that a figure plots the same model described in the Methods. Make sure sample sizes and category labels remain consistent across outputs.
If AI generated multiple representations separately, this cross-check becomes especially important.
Check figures against the underlying image or data integrity requirements
For scientific images, verification may extend beyond ordinary charts. Microscopy images, gels, blots, medical images, and other image-based research data require preservation of data integrity.
Nature's current research-figure guidance states that final images should correctly represent the original data and warns against manipulations that obscure image information. It also prohibits generative AI in figures under its own policy. Other journals and publishers may have different rules, so researchers should verify the policy governing their intended venue.
Do not assume that because an AI tool can modify a research image, a journal permits that modification.
Verify the code if code generated the display
If AI wrote R, Python, MATLAB, or other code to create the table or figure, inspect and test that code as well.
The visualization can faithfully represent the dataset produced by a script while the script itself filtered the wrong observations or calculated the wrong statistic. In that situation, the graphical layer is correct and the analytical pipeline beneath it is not.
Use the dedicated workflow for verifying AI-generated research code when generated code contributes to the display.
Verify the interpretation separately from the representation
A perfectly accurate figure can still be interpreted incorrectly.
If AI looks at a scatterplot and says that X causes Y, the visualization does not establish that causal claim. If two confidence intervals overlap, an AI may make an unsupported statement about statistical significance. If one bar is higher than another, the difference may or may not have the inferential meaning attributed to it.
Once the table or figure itself has been verified, separately check any AI interpretation of the results.
Watch Out
Never verify a research figure merely by asking whether it "looks right." Visual plausibility is particularly weak evidence because many incorrect graphs look entirely ordinary. Trace the representation back to the source data and analysis.
