Should Publication Figures Be Regenerated From Verified Data Rather Than Accepted Directly From Generative AI?

Can You Trust a Publication Figure Just Because AI Made It Look Scientific?

You ask a generative AI tool to create a bar chart comparing experimental groups. Within seconds, it produces a polished figure with labeled axes, attractive colors, and error bars. It looks ready for a journal. But are the plotted values actually derived from your dataset? Do the error bars represent the correct statistical measure? Could the figure contain visual details that were never supported by your analysis?

These questions matter because figures in research publications do more than illustrate findings. They communicate evidence. A convincing appearance cannot establish that a figure accurately represents the observations, calculations, or analytical decisions behind it.

Yes, Publication Figures Should Be Grounded in Verified Research Data

Why a Scientifically Convincing Figure Is Not Necessarily a Scientifically Accurate Figure

A generated picture is not the same as a data-derived visualization

Generative AI systems can produce visual outputs through substantially different mechanisms. An image-generation model may synthesize a chart-like picture from a textual description. A coding assistant, by contrast, may write Python or R instructions that a researcher executes against an actual dataset. Both workflows may be described casually as AI-generated figures, but their evidential foundations are very different.

The distinction is not whether AI participated in creating the figure. It is whether the figure can be connected to an inspectable chain of evidence. A graph generated with AI-written code may be scientifically defensible if the code, data, calculations, and final rendering are verified. A graph synthesized as an image from a prompt may contain invented values even when the prompt describes genuine findings.

What can go wrong when AI creates a chart from a description?

Suppose you instruct an image generator to show that one intervention group achieved a higher mean score than a comparison group. The tool may produce a visually persuasive graph without calculating anything from the observations.

Potential problems include:

• Invented numerical values: Bar heights, data labels, or plotted points may not match the dataset.
• Incorrect scales: Axis intervals, starting points, or category positions may distort the apparent magnitude of a difference.
• Unsupported uncertainty: Error bars may be added without corresponding standard deviations, standard errors, or confidence intervals.
• Missing observations: Outliers, individual data points, or groups may be omitted without a defensible analytical reason.
• Misleading visual emphasis: Colors, shapes, or annotations may imply significance, causation, or categorical distinctions that the analysis does not establish.
• Inconsistent statistical reporting: The figure may display values or significance markers that contradict the manuscript's results.

These are possible failure modes rather than errors that occur in every AI-assisted visualization. The risk is particularly pronounced when the output is a synthesized image rather than a rendering of verified numerical information.

What does regenerating a figure from verified data actually mean?

Regeneration means returning to the data and analytical procedure that support the figure, rather than attempting to make a generated picture appear more credible through visual editing.

A defensible process ordinarily involves identifying the appropriate data version, checking how variables and observations were prepared, reproducing the relevant calculations, and rendering the figure using those verified results.

Regeneration does not necessarily require collecting the data again or repeating every stage of the study. It requires reconstructing the figure from a trustworthy and sufficiently documented analytical foundation. Where a figure was originally produced through a verified computational workflow, independently rerunning that workflow may be more appropriate than recreating the image manually.

Not every publication figure represents the same kind of evidence

A bar chart of observed outcomes, a conceptual model, and a microscopy image serve different scientific purposes. Applying the same verification procedure to all three would overlook important differences.

A conceptual diagram does not necessarily need a numerical dataset. Its accuracy depends on whether the depicted relationships are defensible. Conversely, an experimental image must remain an authentic representation of the research observations. Replacing it with a realistic AI-generated substitute would not establish experimental evidence.

Can AI-written Python or R code be used to generate publication figures?

Yes, subject to applicable journal and institutional policies. A researcher might ask a coding assistant to prepare a plotting script using Matplotlib, ggplot2, or another established visualization library. The script can then be reviewed and executed against the researcher's verified dataset.

However, successful code execution is not equivalent to scientific validation. Code can run without errors while selecting the wrong variable, calculating an inappropriate statistic, excluding observations, or applying misleading plotting defaults.

Researchers should inspect the analytical logic and compare important values with independent calculations or validated outputs. Where AI has contributed to the analysis or figure-generation workflow, the relevant methods and disclosures should make that contribution understandable.

This distinction resembles the broader responsibility to verify AI-assisted reporting of research results. Whether the output is a sentence or a graph, the reported evidence must originate from the study rather than from the model's plausible reconstruction of it.

What do publishers actually allow?

Publisher policies are not uniform, and they may distinguish explanatory illustrations, data visualizations, primary research images, and graphical abstracts.

Elsevier's updated journal policy permits AI-supported explanatory figures under specified conditions. It also permits AI-assisted data visualizations when they are directly derived from underlying data using reproducible analytical or computational methods. Authors must verify the outputs and provide appropriate disclosures. Fabricated data or visual representations that are not faithfully derived from the research are not permitted.

Springer Nature's AI manuscript-preparation policy similarly addresses visual content derived from verifiable inputs, requiring transparency, inspectable source material, human validation, and accurate representation. Individual journals and figure-specific instructions may impose additional restrictions.

Nature's research figure guidance emphasizes faithful representation of original data and prohibits generative AI in figures under that guidance, including content-aware editing of image data. Consequently, a workflow acceptable under one publication's policy should not automatically be assumed acceptable elsewhere.

Does regenerating a figure eliminate every research-integrity concern?

No. Regeneration can establish that the visual output corresponds to a particular dataset and procedure, but it cannot independently establish that the underlying observations were collected properly or that the chosen analysis was appropriate.

A perfectly reproducible chart can still mislead if the dataset contains unaddressed errors, the statistical model is unsuitable, or the figure presents an unjustified interpretation. Researchers therefore need to distinguish computational reproducibility from the broader validity of the research.

Figure captions and manuscript explanations also matter. An accurate chart accompanied by an exaggerated claim can still misrepresent the evidence. The same caution applies when AI-assisted editing strengthens scientific claims beyond what the findings support.

When an Attractive AI-Generated Chart Misrepresents the Actual Results

The essential correction is not simply changing two labels on the AI-generated picture. The researcher replaces an unverified visual construction with a figure derived from the study's actual analytical results.

This example also illustrates why checking AI-assisted scientific reporting against the original evidence matters across both written and visual outputs. A figure and its accompanying results narrative must describe the same analysis.

Common Mistakes When Preparing AI-Assisted Publication Figures

How to Decide Whether an AI-Assisted Figure Is Suitable for Publication

The most useful question is not simply whether AI created the figure. Ask what the figure claims to represent, whether its source can be verified, and whether the intended journal permits the production method.

For researchers preparing manuscripts, a practical approach is to separate visual planning from evidential production. AI may help explore possible layouts or generate draft plotting code, but the publication figure should ultimately be derived from a documented source and reviewed against the analysis.

Preserve the final figure together with the relevant source data, scripts, and figure specifications. This makes later corrections easier and helps when coauthors, reviewers, or editors ask how a visualization was produced.

If AI was used materially in figure generation or analytical processing, document its role and follow the journal's disclosure instructions. The distinction between assistance and scientific contribution also matters when describing the methods actually used in the study.

What to Verify Before Submitting an AI-Assisted Research Figure

Frequently Asked Questions About AI-Generated Publication Figures

Publication Figures Must Represent Evidence, Not Just Resemble It

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.