03 · What You Need to Know
How AI-Generated Tables and Figures Can Misrepresent Research Evidence
What Counts as a Fabricated Research Table or Figure?
A fabricated research table or figure presents invented or materially misrepresented information as though it originated from genuine research observations or analyses.
Examples may include a demographic table describing participants who were never recruited, a statistical results table containing invented coefficients, or a scientific image presented as an experimental observation that was never obtained.
However, artificial content is not inherently fraudulent.
Researchers may legitimately use simulated datasets, schematic diagrams, or illustrative charts when their origin and purpose are accurately described.
The distinction concerns whether the material is represented honestly and whether readers can understand its evidentiary status.
Synthetic Data and Fabricated Empirical Evidence Are Different
Synthetic data are artificially generated observations that may be used in simulation, software testing, methodological research, or teaching.
For example, a statistics instructor may create a hypothetical dataset to demonstrate regression analysis.
Such use is legitimate when the values are identified as simulated and conclusions are interpreted within the simulation framework.
Fabrication becomes a concern when synthetic observations are falsely described as data collected from actual participants or experiments.
Legitimate synthetic illustration
A table, figure, or dataset explicitly identified as simulated, hypothetical, or illustrative, with its purpose and limitations accurately described.
Reporting or visualization error
An inaccurate table or figure resulting from incorrect calculations, plotting, labeling, or other mistakes requiring correction.
Potential fabrication or falsification
Invented or materially altered research evidence represented as genuine findings, subject to the applicable misconduct definition and evidentiary requirements.
The origin and representation of the evidence matter more than whether the values appear realistic.
What Types of Research Tables Can AI Generate?
Generative AI can produce many familiar table structures used in academic manuscripts.
| Table type |
Potential AI-generated content |
Verification requirement |
| Demographic table |
Participant frequencies, percentages, ages, and group characteristics |
Compare with actual participant records and data summaries |
| Descriptive statistics |
Means, standard deviations, medians, and ranges |
Recalculate from the underlying dataset |
| Hypothesis testing |
Test statistics, p-values, and confidence intervals |
Verify statistical procedures and calculations |
| Regression results |
Coefficients, standard errors, and model fit statistics |
Reproduce the fitted model and examine assumptions |
| Psychometric results |
Reliability coefficients, factor loadings, and validity-related statistics |
Verify against instrument data and analytical outputs |
| Qualitative findings |
Themes, participant quotations, and coding frequencies |
Trace findings to authentic research material and analytical documentation |
AI may help format genuine results into these structures. However, it should not invent values to fill missing cells or create the appearance of completed analyses.
Can AI Invent Descriptive Statistics?
Generative AI can produce plausible means, standard deviations, percentages, and frequency distributions.
For example, it might generate a demographic table reporting that 60% of participants were female and 40% were male, even when no participant data were supplied.
These percentages may look reasonable but have no empirical basis.
Researchers should calculate descriptive statistics from actual data using appropriate analytical procedures.
AI-generated summaries should be checked against independently reproducible calculations.
A plausible value is not evidence that the corresponding observation exists.
Can AI Fabricate P-Values and Confidence Intervals?
Generative AI can produce statistical results that resemble outputs from established analytical software.
However, generated p-values and confidence intervals may not correspond to the reported sample size, effect estimate, standard error, or statistical test.
For example, a manuscript might report a regression coefficient and confidence interval that are inconsistent with the stated standard error.
Such discrepancies may be detected through mathematical verification.
However, even mathematically consistent statistics can be invented.
Researchers should verify results against the actual analysis and underlying data rather than rely on internal numerical plausibility alone.
Can AI Generate Plausible Regression Tables?
Regression tables commonly report coefficients, standard errors, test statistics, confidence intervals, and model fit measures.
Generative AI can reproduce this format and populate it with plausible numerical values.
However, the resulting table may not correspond to any fitted model.
Researchers should verify variable definitions, analytical samples, model specifications, estimation procedures, and reported coefficients.
They should also confirm that the selected model is scientifically appropriate.
The existence of a complete regression table does not establish that the analysis was performed or that the model supports the conclusions.
Can AI Fabricate Qualitative Research Findings?
Qualitative research may present themes, categories, participant quotations, and interpretive summaries.
Generative AI can produce plausible quotations and thematic narratives without access to actual interview transcripts or observation records.
For example, it may generate a quotation attributed to a fictional participant describing positive experiences with online learning.
Presenting that quotation as genuine participant testimony would misrepresent the research evidence.
Researchers should ensure that quotations and interpretations are grounded in authentic data and the analytical approach used.
Confidentiality may limit public access to transcripts, but appropriate research records should support the reported findings under applicable requirements.
What Research Figures Can Generative AI Produce?
AI systems can generate charts, diagrams, illustrations, and image-like material.
Different outputs have different evidentiary implications.
| Figure type |
Legitimate use |
Potential integrity problem |
| Bar or line chart |
Visualizing genuine data or clearly labeled simulations |
Displaying invented observations as empirical results |
| Scatterplot |
Representing observed relationships or simulated examples |
Fabricating data points to suggest a desired association |
| Conceptual diagram |
Illustrating theories, models, or proposed mechanisms |
Presenting an unsupported relationship as empirically established |
| Scientific illustration |
Explaining structures or procedures with appropriate labeling |
Misrepresenting an illustration as an actual observation |
| Microscopy-like image |
Clearly labeled educational or synthetic material |
Presenting generated imagery as experimental micrographs |
| Medical or diagnostic image |
Authorized simulation or illustration |
Representing generated content as genuine patient or research imaging |
Readers should be able to distinguish an illustration from an observation and a simulated chart from an empirical result.
Why Can Fabricated Charts Be Particularly Persuasive?
Charts communicate patterns visually, often more immediately than numerical tables.
A bar chart showing a large treatment effect may create a strong impression of effectiveness even when the underlying values are invented.
Similarly, a scatterplot may suggest a compelling relationship through the placement of points.
Readers may focus on the apparent trend without examining the provenance of the observations.
Researchers should therefore treat visualizations as representations of evidence rather than independent evidence of authenticity.
The scientific validity of a chart depends on the data and analytical choices that produced it.
AI Can Also Produce Misleading Figures From Genuine Data
Not every misleading figure contains fabricated observations.
A chart may use genuine data but misrepresent the findings through inappropriate axis scaling, omitted uncertainty, selective time windows, or incorrect labels.
For example, a bar chart with a truncated vertical axis may exaggerate a small difference between groups.
Similarly, a figure may omit important comparison groups or display percentages without identifying their denominators.
These problems may arise from mistakes, poor design choices, or deliberate manipulation.
Researchers should verify both the underlying values and the visual encoding used to communicate them.
Can AI Alter Scientific Images?
Generative image tools can create or modify visual content, including material that resembles microscopy, medical imaging, or laboratory observations.
Such capabilities may be useful for clearly labeled illustrations and educational demonstrations.
However, generated or altered imagery must not be misrepresented as an authentic experimental observation.
Scientific image integrity requires attention to original acquisition files, processing procedures, and relevant disciplinary standards.
The principles governing AI detection of falsified scientific images remain relevant when evaluating possible manipulation.
Image appearance alone may not establish whether the content is authentic.
Can AI Generate Figures That Evade Visual Detection?
Generated scientific images may resemble genuine observations according to some visual characteristics.
However, the ability to produce realistic imagery does not establish that every generated image will evade forensic analysis.
Detection performance depends on the image-generation method, acquisition context, processing history, and available reference evidence.
Some generated images may contain visible or statistical anomalies, while others may not.
Researchers should avoid assuming that a detector can reliably distinguish every synthetic image from an authentic scientific observation.
Original data provenance is often more informative than visual appearance alone.
What Is Data Provenance in Research Tables and Figures?
Data provenance concerns the origin of research information and the procedures through which it was transformed into reported outputs.
For a statistical table, provenance may include the source dataset, variable definitions, analytical code, software environment, and calculation outputs.
For a scientific image, it may include acquisition files, instrument settings, processing records, and figure assembly documentation.
For a qualitative table, it may involve the relationship between authentic source material, coding decisions, and reported themes.
Provenance supports verification by allowing reviewers to understand how a reported result was produced.
However, the existence of a file or script does not automatically establish that the underlying observations are genuine.
How Can Researchers Verify AI-Generated Tables?
Researchers should compare each consequential value with the output of an appropriate analysis performed on the actual dataset.
For example, demographic frequencies should reconcile with the analytical sample, and percentages should correspond to the stated denominators.
Regression coefficients should match the fitted model and its documented specification.
Where possible, tables should be generated directly from reproducible statistical scripts rather than manually populated from conversational AI output.
AI may assist with formatting, but its changes should not alter the underlying numerical information.
Researchers should preserve a clear distinction between verified calculations and generated explanatory text.
How Can Researchers Verify AI-Generated Figures?
Verification should begin with the underlying source data and the intended scientific message.
Researchers should check that plotted values correspond to the dataset, that axis labels and units are correct, and that uncertainty is represented appropriately.
For scientific images, the original acquisition and processing history may require examination.
Where AI generates a conceptual or illustrative figure, the caption should make its nature clear when relevant.
Reviewers should also assess whether the figure communicates relationships that the evidence can actually support.
Visual accuracy and substantive interpretation are both important.
Can Statistical Consistency Checks Detect Fabricated Tables?
Some table errors can be identified through arithmetic or statistical verification.
For example, percentages may fail to correspond to reported frequencies, or confidence intervals may be incompatible with the stated estimates under the relevant method.
Researchers can also examine whether reported means are possible given the measurement scale and sample size.
However, these checks depend on assumptions and may not apply to weighted or adjusted estimates.
The limitations of AI statistical anomaly detection should guide their interpretation.
A mathematically consistent table can still contain invented values, while a genuine table can contain correctable errors.
Can AI Fabricate Tables That Support a Preferred Hypothesis?
Generative AI can produce numerical examples illustrating a specified relationship.
Such examples may be legitimate in teaching or simulation when clearly labeled.
However, selecting invented values to make a hypothesis appear empirically supported would misrepresent the research.
Researchers should not ask AI to fill missing empirical results with plausible numbers or adjust values until they produce a desired conclusion.
The broader question of whether AI should determine hypothesis support reinforces the need to base scientific conclusions on authentic evidence rather than preferred outcomes.
What If AI Changes Numbers While Formatting a Table?
AI systems may introduce numerical errors even when their intended role is limited to formatting.
For example, a system may round values inconsistently, transpose group labels, omit rows, or generate an incorrect total.
Such errors may occur without deliberate fabrication.
Researchers should compare the final table with the verified source output.
For consequential results, automated consistency checks or independent review may be appropriate.
The research team remains responsible for the accuracy of the published table.
Can AI-Generated Figures Be Used in Legitimate Research Publications?
AI-generated illustrations may be appropriate when they serve a clearly identified explanatory purpose and comply with applicable publication policies.
For example, a conceptual diagram may help communicate a proposed theoretical framework.
However, policies concerning generative images vary among publishers and journals.
Some outlets restrict or prohibit particular uses, especially when generated imagery could be mistaken for primary research evidence.
Researchers should verify the applicable requirements before submission and disclose AI involvement where required.
Permitted illustration does not authorize the replacement of genuine experimental evidence with synthetic images.
How Are Fabricated Tables and Figures Related to Paper Mills?
Paper-mill manuscripts may contain invented numerical results, manipulated scientific images, or repeated visual material.
Generative AI may reduce the effort required to produce plausible supporting content.
However, a suspicious table or figure does not independently establish organized paper-mill activity.
The broader issue of generative AI and academic paper mills requires evidence about research authenticity and publication-process manipulation.
Researchers should evaluate the specific discrepancy rather than infer the origin of an entire manuscript from one visual anomaly.
Why Detection Alone Cannot Establish Fabrication
AI-assisted screening may identify unusual numerical distributions, duplicated images, or inconsistencies between tables and text.
These observations can justify verification.
However, statistical anomalies may arise from legitimate data-generating processes, while visual similarities may reflect appropriate reuse.
Conversely, invented material may appear statistically and visually plausible.
The distinction between AI suspicion and evidence of research misconduct is therefore essential.
A detector's output cannot independently establish the origin or intent behind the material.
What Should Editors Examine When Figures Appear Fabricated?
Editors should identify the specific concern and seek appropriate evidence.
For statistical tables, this may involve verifying calculations or requesting clarification of analytical procedures.
For scientific images, it may involve original files and relevant processing documentation.
For charts, the underlying data and visualization code may be informative.
Review should be proportionate and respect confidentiality, privacy, and institutional requirements.
The absence of a particular record may limit verification without automatically establishing misconduct.
Research Misconduct Findings Require More Than an Incorrect Figure
A table or figure may be incorrect because of an honest mistake, inappropriate analytical choice, or deliberate misrepresentation.
These possibilities should not be treated as equivalent.
Under the United States Public Health Service framework, formal research misconduct findings require specified evidentiary and culpability elements.
Editors may need to correct or reject unreliable material even when misconduct has not been established.
The appropriate response depends on the verified evidence and applicable procedures.
The principles governing human review of AI research-integrity flags apply to both numerical and visual concerns.
How Can Research Teams Prevent AI-Related Table and Figure Errors?
Researchers should maintain a reproducible connection between the source evidence and reported outputs.
Where feasible, statistical tables and charts should be generated directly from validated analytical workflows.
Scientific images should retain appropriate original files and documented processing histories.
AI-assisted formatting or illustration should be reviewed to ensure that it does not introduce unsupported values, labels, or visual elements.
Researchers should also distinguish empirical results from hypothetical examples and disclose relevant AI assistance under applicable policies.
The objective is to preserve evidentiary integrity throughout the reporting process.
Watch Out
Do not use AI to fill missing empirical tables with plausible values, invent statistical results, or generate images that are presented as actual experimental observations. A convincing table or figure is not evidence unless its content can be traced to an appropriate source.
