In June 2023, Nature told authors something that would have sounded odd five years earlier: images and video made with generative AI were no longer acceptable in its journals. The announcement drew a line that had been missing. Text produced by a large language model could sometimes be allowed with careful disclosure, but pixels produced by a diffusion model were treated as a different risk. Since then, the Science family of journals, the JAMA Network and the broader Springer Nature portfolio have moved in the same direction, though the details vary enough to catch authors off guard.
What used to be a metadata question — who made this figure, and with what tool — has become an editorial compliance question. Journals are not rejecting AI because the tools are new. They are rejecting images because images do not leave a reliable audit trail. A text passage can be checked against prompts, drafts and model logs. A synthetic western blot, a generated microscopy panel, or an AI-altered graph where the error bars were invented rather than measured cannot be checked the same way.
The new split between text and pixels
Nature's 2023 policy did something unusually useful: it forced authors to separate the rules for text, code, raw data and images instead of treating them as one problem. The result is a split that now appears in submission checklists. An author who uses a language model to polish prose may pass with a declaration. The same author with an AI-generated schematic runs into an outright prohibition.
Science journals take a similar line, reserving AI-generated figures for cases that receive explicit editorial approval and requiring authors to disclose any use of generative AI anywhere in the work. JAMA Network has taken a similar stance: AI-generated images are considered only when they are part of the research itself. Many Elsevier and Taylor & Francis journals now require disclosure of AI-assisted figure generation and ask editors to request original source files when an image's provenance is unclear.
Why images got singled out
Image manipulation is not a generative AI problem. It is a digital biology problem that predates diffusion models by a decade. Journals began deploying forensic screening after duplicated bands and cloned microscopy fields became the most common reason for image-level retractions. The arrival of AI image generators compressed the timeline: an author with no bench background could produce a plausible cell image in minutes, and traditional source-file checks stopped working. That is why image-manipulation detection tools are now embedded earlier in the review process.
Nature's June 2023 editorial pointed to the core issue in blunt terms. Generative AI imagery cannot be traced to data, methods or source files. That absence of provenance, not aesthetic quality, drives the restriction. The editorial also noted exceptions for research about AI systems, a carve-out that keeps the policy from blocking legitimate work in machine learning and computer vision.
The base rate and the exception
Before generative AI, the base rate for problematic images was already measurable. In a screen of 20,621 papers across 40 journals, Elisabeth Bik and colleagues found inappropriate image duplication in 3.8 percent. That is roughly one paper in 26, and it is the floor, not the ceiling, because the study relied on visual inspection rather than automated detection across every figure.
The exception to the new bans is narrower than many authors expect. Journals that prohibit AI images still allow AI-assisted processing such as denoising or deconvolution and simple contrast adjustment that leaves pixel values intact, when it is disclosed and does not fabricate details. The test is usually one question: did the tool invent something that was not in the underlying data? If yes, it belongs in the forbidden category. If no, it may be acceptable with documentation.
A collaborator — call her Dr. O — used a text-to-image model to make a schematic of an experimental workflow for a review. It looked clean, so she submitted it. The journal asked for source files; there were none. No layers, no vector paths, no script, just a PNG. The figure was rejected on documentation, not on content. She redrew it in BioRender the same afternoon and the revised file passed. The policy did not question her honesty; it rejected evidence that could not be audited.
What authors now have to tell journals
Disclosure requirements have moved from broad acknowledgements to line-item declarations. The most recent ICMJE update on authorship and AI disclosure reinforced that AI cannot be an author and that its use in any part of the work, including figure creation, must be reported. Authors remain accountable for every panel, whether a model drafted it or not.
Science journals ask authors to specify which tool was used and how, and editors can request the original unprocessed data behind any figure. Springer Nature's broader guidance sets out separate declarations for language editing, image processing and code generation. A manuscript that uses a grammar checker for text sits in a different category from one that uses a diffusion model for a figure, and the submission form now asks which one the author did. The Springer Nature AI policy is a useful crib sheet when the journal-specific form is vague.
What this means for your lab
The practical burden falls on labs that produce figures, not on publishers. That burden is small if it is built into the figure pipeline and large if it is discovered during review.
- Keep the raw image file, the processed file and the script or settings that connect them. A PNG alone is no longer a source file.
- If you use AI-assisted processing, record the tool, version, prompt or settings and the specific panels it touched.
- Check the target journal's figure policy before you start drawing, not when you upload. A banned figure discovered at revision costs more than the time to redraw it.
- Assign one person in the lab to own the image documentation folder. That habit converts a compliance request from a scramble into a five-minute email.
None of this makes figures more scientific. It makes them more checkable, which is the property editors say they can enforce at scale.
Where the rules could break
These policies are cleaner on paper than in practice. Disclosure depends on author memory, and detection tools do not reliably label a pixel as synthetic. They find duplication, cloning, inserted filters and compression artifacts, then editors use those findings to request raw files.
There is also a licensing dispute that sits upstream of every disclosure form. Training data for text-to-image models includes figures drawn from journals and preprint servers, and some publishers have signed agreements to license their archives. Authors who object to AI training now face a split between their institution's enthusiasm for AI and their publisher's image policy. The legal fights over publisher AI training will shape what authors can confidently submit years before the detection tools catch up.
Before you submit
Open the journal's editorial policy and search for the word 'image.' If the policy bans AI-generated figures, plan to supply source files for every panel. If it allows them with disclosure, write the exact tool and version in the cover letter and the methods. Do it before submission, not after the editor asks.
Photo by Brett Jordan on Unsplash