When a published figure turns out to be AI-generated, the question that follows is never “which tool made this.” It is who signed off on it. Across every major publisher policy, the answer is the same: the human authors on the byline, not the AI tool, not the person who happened to run the prompt. But publishers do not treat all AI-generated visual content the same way. A figure that purports to show real experimental data is judged very differently from a schematic diagram, flowchart, or graphical abstract that illustrates a concept. This guide covers both halves of that question: where accountability sits, and where the permitted/prohibited line actually falls.
This page focuses specifically on figures and images. For the broader landscape of how publishers handle generative AI in manuscript text, peer review, and editorial workflows, see Journal and Publisher Policies on Generative AI in Manuscripts.
The core rule: authors are accountable for every figure, regardless of how it was made
Every major publisher policy converges on one point before it gets to any distinction between figure types: an AI tool cannot be an author, cannot hold a conflict of interest, cannot sign a copyright agreement, and cannot be held accountable for the work. That reasoning is spelled out explicitly in the ICMJE generative AI guidance (added to the ICMJE Recommendations in the May 2023 update) and in COPE’s position statement on authorship and AI tools (published February 2023), both of which state that AI tools do not meet authorship criteria because they cannot take responsibility for the submitted work.
That non-authorship of the tool is precisely what makes the human authors accountable by default. ICMJE’s own fourth authorship criterion, part of the standard test covered in authorship criteria content elsewhere on this site, requires every author to agree to be accountable for all aspects of the work, including its accuracy and integrity. A figure generated or altered with an AI tool does not create an exception to that criterion; if anything, it raises the bar, because the corresponding author (and, per most policies, every co-author) is expected to be able to explain how a given figure was produced and to vouch that it accurately represents what it claims to represent.
Elsevier states this directly in its author-facing generative AI policy: authors must verify the accuracy and originality of all images submitted, ensure images reflect the authors’ own work, and remain fully responsible for the content and editorial process regardless of whether AI tools were involved in creating or altering any part of it. IEEE’s guidance is structurally identical: human authors are 100% accountable for a submission regardless of AI involvement, and AI-driven data or figure fabrication is treated as a research-integrity violation, not a disclosure technicality — see the IEEE generative AI policy term and the related comparison of IEEE’s AI policy against its 3-part authorship test for how that connects to IEEE’s underlying authorship criteria.
Two categories of AI-generated visual content — and why publishers split them
Publisher policies on AI-generated images almost universally draw a line between two categories, even though the exact wording differs journal to journal:
- Data figures and images that purport to represent actual research results — micrographs, histology and pathology images, western blots, radiology or imaging scans, gel images, photographs of specimens, and any chart or plot generated from real experimental measurements. Generating or materially altering these with AI is largely prohibited, because an AI-generated data image can misrepresent findings that never actually occurred, which is a research-integrity problem, not a formatting one.
- Schematic diagrams, illustrations, flowcharts, and graphical abstracts — explanatory or conceptual visuals that do not claim to be a record of an experimental result. These are generally permitted with disclosure, on the reasoning that they illustrate an idea or a workflow rather than asserting a specific finding, so the integrity risk of a wrong pixel is much lower than it is for a data image.
The dividing question every policy is effectively asking is: does this image assert that “this is what we observed,” or does it assert “this is how the process works”? The first kind carries the risk of scientific misinformation if the AI fabricates or distorts detail; the second does not, provided it is disclosed and does not misrepresent the underlying science it is illustrating.
That said, “generally permitted” for the second category is not the same as “unrestricted.” Several publishers narrow even the illustrative-image exception to AI features built into dedicated scientific-illustration tools (such as BioRender), rather than open-ended, general-purpose image generators — see the publisher-by-publisher detail below. A schematic is only in the permitted category if it stays schematic: the moment an illustration starts standing in for actual data (for example, an AI-rendered “representative” microscopy image used in place of a real one), it has crossed into the prohibited category regardless of what the figure is labeled.
How specific publishers draw the line
Elsevier
Elsevier’s generative AI policy for journals states that AI tools must not be used to create or alter images that represent primary observed or experimental data (its examples include microscopy, histology, western blots, radiology scans, and patient images), and that AI must not be used to fabricate results, invent or alter underlying data, or generate figures not faithfully derived from the actual data and methods used. Explanatory images — flow charts, conceptual diagrams, decision trees — are permitted, with disclosure in the figure caption and a general AI-use statement. Elsevier narrows the graphical-abstract case specifically: authors are expected to use dedicated scientific-illustration tools rather than general-purpose AI image generators for graphical abstracts. Where AI-assisted imaging is genuinely part of the research method itself (for example, AI-assisted image reconstruction in biomedical imaging), that use must be described reproducibly in the Methods section, naming the tool, version, and developer. Elsevier’s separate graphical abstract format and size specifications are covered in this site’s graphical abstract guide.
Nature Portfolio
Nature Portfolio’s editorial position, set out in its 2023 editorial “Why Nature will not allow the use of generative AI in images and video,” is that the journals are unable to permit AI-generated images or video in their content, citing unresolved legal and integrity issues, with only narrow, editor-reviewed exceptions such as content where the AI-generated image is itself the subject of the piece being discussed. Non-generative AI or ML tools used to manipulate or enhance an existing figure are treated as a separate, lighter-touch category requiring caption disclosure and case-by-case editorial review rather than an outright ban. Nature’s stance is notably stricter on images specifically than several competitor policies. See the Nature Portfolio AI Policy entry for the full picture across text and images.
Science (AAAS family)
Science’s editorial policies, applied across Science, Science Advances, Science Immunology, Science Robotics, Science Signaling, and Science Translational Medicine, do not allow the use of AI to create figures, and require authors to certify and take responsibility for all content submitted, including anything generated with AI assistance. Undisclosed AI-generated image use is treated as a form of scientific misconduct, one of the firmer enforcement stances among the major journal families. Since 2024, the Science family has also used Proofig, an AI-assisted image-analysis tool, to screen submitted images across all six journals for signs of duplication or manipulation — a screening approach also adopted elsewhere, as covered in this site’s news coverage of AI image-integrity screening adoption at MDPI and ASM.
Cell Press
Cell Press’s graphical abstract guidelines take the same “dedicated tool, not general-purpose generator” approach as Elsevier (Cell Press is an Elsevier imprint): general-purpose generative AI image tools are not permitted for graphical abstract creation, while AI features built into dedicated scientific-illustration software are allowed.
IEEE
IEEE does not centralize its generative AI policy in a single branded document; guidance is issued through the IEEE Author Center and echoed, with consistent substance, by individual IEEE technical societies. The consistent thread across that guidance is the accountability framing above: AI tools cannot be listed as co-authors, disclosure of AI use is typically expected in the acknowledgments, and AI-driven data or figure fabrication is treated as a research-integrity violation regardless of how the fabrication happened.
MDPI
MDPI’s generative AI/LLM authorship policy requires disclosure of substantive AI-generated content (naming the tool in Acknowledgments, describing use in the Materials and Methods section or equivalent) and states plainly that authors remain fully accountable for all AI-assisted content, explicitly including data and image manipulation. Basic language editing is exempt from disclosure; image and data generation are not.
JAMA Network / AMA Manual of Style
JAMA Network’s guidance (Flanagin, Kendall-Taylor, and Bibbins-Domingo, JAMA, July 2023) follows the same shape as the other policies above: AI tools cannot be listed as authors, AI-generated content requires disclosure naming the tool, version, and manufacturer, and authors must confirm responsibility for the accuracy of anything AI-assisted, images included.
Why the split exists: the integrity rationale
The prohibition on AI-generated data figures is not a formatting preference; it is a direct extension of ordinary image-manipulation and image-integrity norms that predate generative AI by decades. A western blot, a microscopy image, or a scan is treated by editors and by research-integrity offices as a form of primary data: altering it to show something that was not actually observed is scientific misconduct whether the alteration was done with a clone-stamp tool in Photoshop or with a diffusion model. Generative AI simply makes that kind of alteration faster, more convincing, and harder to detect by eye — which is exactly why journals have paired policy bans with technical screening tools (Proofig, ImageTwin, and similar image-forensics software) rather than relying on disclosure alone.
Schematic diagrams and graphical abstracts do not carry that same risk profile, because they are not offered as evidence of what was observed. An AI-generated flowchart that is slightly stylistically odd is a cosmetic problem; an AI-generated “representative” data image that never came from an actual sample is a fabrication problem. That is the underlying logic every publisher policy above is applying, even where the exact wording differs.
What accountability looks like in practice for authors
Being “accountable” for a figure is not a passive disclosure exercise; publisher policies describe several concrete obligations that fall on authors, typically the corresponding author on behalf of the full author group:
- Know how every figure in the manuscript was actually produced. If a co-author generated or edited a figure using an AI tool, the corresponding author is expected to be able to explain that to an editor or reviewer, not merely take it on faith.
- Disclose AI use for figures separately from AI use for text. Most policies expect image-specific disclosure — typically in the figure caption, a dedicated AI-use statement, or the Methods section — distinct from any general statement about AI-assisted writing. A single blanket “AI was used to help write this paper” statement does not satisfy an image-specific disclosure requirement.
- Name the tool. Policies that require disclosure (Elsevier, MDPI, JAMA, IEEE) consistently ask for the specific tool name, and often version and developer, not just “generative AI” as a category.
- Be prepared for image-integrity screening. Automated tools now routinely check submitted figures for duplication, splicing, and AI-generation signatures before or during peer review at several major publishers.
- Understand that disclosure does not cure a prohibited use. Disclosing that a data figure was AI-generated does not make it publishable at a journal that prohibits AI-generated data figures outright — disclosure is a requirement layered on top of what is permitted, not a workaround for what is not.
This dovetails with how AI-assisted contributions are represented in a CRediT contribution statement where one is used: see How to Disclose AI Assistance in a CRediT Statement, Role by Role and CRediT for Non-Traditional Contributors: Community and AI-Assisted Work for how figure-generation assistance specifically maps onto roles such as Visualization or Software.
Frequently asked questions
Can an AI tool be credited as a co-author for generating a figure?
No. Every major publisher and standards body position covered above — ICMJE, COPE, Elsevier, Nature Portfolio, Science, IEEE, MDPI, JAMA — agrees that an AI tool cannot be listed as an author or co-author, because authorship requires the ability to take responsibility for the work, hold a conflict of interest, and agree to publication terms, none of which a tool can do.
Is a graphical abstract held to the same standard as a data figure?
Generally not, but it is not unrestricted either. Most publishers permit AI assistance for graphical abstracts with disclosure, though several (Elsevier, Cell Press) specifically expect that assistance to come from dedicated scientific-illustration tools rather than general-purpose AI image generators, and none of the permissive policies extend to a graphical abstract that misrepresents the actual findings it summarizes.
What happens if an AI-generated data figure is discovered after publication?
Publisher policies treat this as a research-integrity matter handled through standard post-publication channels — correction, expression of concern, or retraction depending on severity — the same pathway used for other undisclosed image manipulation, rather than as a separate “AI” track. Authors remain accountable regardless of how much time has passed since publication.
Does using an AI tool to enhance an existing figure (not generate a new one) count the same way?
No — most policies, including Nature Portfolio’s, treat non-generative AI or ML-based enhancement of an existing, real image (contrast adjustment, denoising, resolution upscaling) as a separate, lighter-touch category from generating a new image from a prompt. It generally still requires disclosure, but is reviewed case-by-case rather than banned outright, provided the underlying image is genuine and the enhancement does not add or remove content that changes the scientific meaning.
Where should AI-generated-figure disclosure go in a manuscript?
Conventions vary by publisher: commonly the figure caption itself, a dedicated AI-use disclosure statement (often near the Acknowledgments), and/or the Methods section when the AI tool was part of the actual research method rather than just figure preparation. Check the specific journal’s author guidelines, since placement is one of the more variable elements across otherwise similar policies.
