AI-Generated Images in Journals: Disclosure Rules

Short answer

Publishers differ, but the pattern is consistent. Images that present research results are restricted or barred from generative AI creation or alteration, illustrations are allowed on conditions, and permitted use must be disclosed, naming the tool and purpose, in the caption, an AI declaration or the methods. Read the target journal’s current policy first.

This guide is for authors, journal editors, peer reviewers and research integrity officers. After reading it you should be able to classify an image, find the disclosure each publisher asks for, write a complete statement and check a claim. Policy details below were read from each publisher’s own page on 2026-10-30 and change over time, so re-check before you submit.

How this guide differs from neighbouring CASRAI pages: it covers image-specific publisher standards and provenance labelling. It is not a general guide to AI text disclosure (see AI disclosure statement examples) and not a forensic technique guide (see detecting image manipulation in figures). For who is answerable for a figure, see author accountability for AI-generated figures.

What counts as an AI-generated or AI-altered image?

An image is AI-generated when a generative model produced it from a prompt or other input, and AI-altered when a model added, removed or changed content in an image that existed before. Routine adjustment of a captured image is a different category, and the policies treat it differently. The dividing question is whether the image claims to show something that was observed.

Research-data images versus illustrations

Two categories recur across the policies read for this guide.

Research-data images are pictures, plots or scans that stand for an observation or measurement: micrographs, blots, gels, clinical or diagnostic images, photographs of samples, charts built from real measurements. Elsevier, for example, does not permit generative AI or AI-assisted tools to create or alter images in submitted manuscripts, and Taylor and Francis does not permit generative AI in the creation or manipulation of outputs of research, naming research results, clinical samples and diagnostic images.
Illustrations are conceptual figures, flow diagrams, schematics and teaching images that do not claim to be a record of an experiment. Taylor and Francis permits generative AI to assist research data visualisations and conceptual illustrations of processes; Wiley says AI may be used for creating data visualizations and illustrations; PLOS accepts AI-generated figures provided they are not used to misrepresent primary research outputs.

Acceptable routine adjustment

Ordinary adjustment is not the same thing as AI alteration. Elsevier states that brightness, contrast or colour balance adjustments are acceptable if they do not obscure or eliminate any information. Wiley notes that some edits are legitimate, such as enlarging part of an artwork, provided the integrity of the underlying data is not altered, and asks authors to keep original, unprocessed images. Taylor and Francis defines manipulation to include augmenting, concealing, moving, removing or introducing a specific feature. In practice, the question for an editor is whether the edit could change what a reader concludes from the image.

Springer Nature adds a third distinction. Its policy defines visual content broadly (figures, graphs, charts, maps, diagrams, illustrations, photographs, images and video) and separates generative AI from non-generative tools used to edit or analyse existing visuals, with different conditions for each.

What do publishers and bodies require? A policy comparison

The table summarises what each source says about images, read from its own page. “No date shown” means the page displayed no policy date when checked. Each row states only what that source’s own page says.

Source	What it says about AI-generated or AI-altered images	Disclosure it asks for	Date shown on page
Springer Nature, AI use in manuscript preparation	AI may create, modify, enhance or analyse visual content where use is transparent and lawful. Generative AI is acceptable only if authors meet conditions: disclosure with model and purpose, inspectable and attributed source material, human review and validation, an output that accurately represents the evidence, and no infringement of copyright or licensing. Non-generative tools are acceptable where the underlying material remains attributable and changes do not misrepresent it. Content without verifiable sources is described as opaque and not permitted.	An AI Declaration in the manuscript and a note in the caption or legend, including model and purpose	No date shown
Elsevier, generative AI in writing	Does not permit generative AI or AI-assisted tools to create or alter images in submitted manuscripts, and not for artwork such as covers. Exception: where such use is part of the research design or methods, described in a reproducible manner in the methods section. Authors may be asked for pre-AI-adjusted versions.	A separate AI declaration statement at submission, placed above the references; research-process AI use goes in the methods	“Policy updated: July 2026”
Wiley, ethics guidelines	AI may be used for data visualizations and illustrations, with important restrictions on photograph enhancement. Inappropriate image manipulation is considered data fabrication or falsification. Authors should keep original, unprocessed images and provide them on request.	Disclosure when AI substantially edits, develops or translates a manuscript, described transparently; the journal editor decides whether use is appropriate	“Last updated: July 2026”
Taylor and Francis, AI policy	Permits generative AI to assist research data visualisations and conceptual illustrations; does not permit generative AI in creating or manipulating outputs of research or clinical testing (research results, clinical samples, diagnostic images). Points to a separate journals policy on images and figures.	Clear acknowledgement in the article with the full tool name and version, how it was used and why; a specific disclosure statement for articles	No date shown
SAGE, ChatGPT and generative AI	Lists generated images presented as unique or novel research images as an inappropriate use. Data visualization assistance and representative illustrations or infographics are uses that require disclosure. Authors must verify output accuracy.	At submission, in the methods or acknowledgements; citing the tool may be appropriate for images and visualizations	No date shown
PLOS, ethical publishing practice	AI-generated figures allowed provided AI content is not used to misrepresent primary research outputs. Images must not be manipulated in any way that could lead to misinterpretation. Authors must check accuracy and validity of AI-generated figure content.	Disclosure in the figure legend; disclosures name the tools, describe how they were used and explain how validity was checked	No date shown
ICMJE, AI use by authors	Does not discuss figures separately but names image creators among AI-assisted technologies. Authors must be able to assert there is no plagiarism, including in text and images produced by the AI. Humans are responsible for submitted material.	Indicate at submission whether AI-assisted technologies were used and describe the use in the cover letter and the submitted work where relevant	No date shown

Two cautions. First, journal-level instructions can be narrower than a publisher-wide page, and journals may add their own rules. Second, the table records what the pages say, not how strictly each rule is enforced. For the fuller write-ups of two publishers, see CASRAI’s guides to the SAGE generative AI policy and the Taylor and Francis generative AI policy.

How do you disclose an AI-generated image in a paper?

Disclose in every place the target journal names, which across the pages above means some combination of the figure caption or legend, a dedicated AI declaration, the methods section and the cover letter. A single sentence about AI-assisted writing does not cover an image. The elements below combine what the publisher pages ask for with editorial suggestions; the marked items are not required by the pages read.

Tool name and version. Taylor and Francis asks for the full tool name with version number; Springer Nature asks for the model.
Which figure or panel. Identify each affected image by number so reviewers do not have to guess (editorial suggestion; Springer Nature and PLOS place the disclosure in the caption or legend, which does this automatically).
Role of the tool. State what the tool did: created the image, edited a captured image, enhanced or analysed it. Springer Nature distinguishes generative from non-generative use, and Taylor and Francis asks how it was used and why.
Date of use, and prompt or settings. Not required by the pages read. Recording them is an editorial suggestion that makes the use reproducible and answers an editor’s follow-up question; Elsevier requires a reproducible description where AI is part of the research methods.
Human review. Springer Nature expects human review and validation, and PLOS expects authors to say how they checked validity.
Source material and rights. Springer Nature requires that source material be verifiable and attributed and that use not infringe copyright or licensing. Whether a given tool’s terms or training data allow the intended publication is an author responsibility to check; the pages read do not settle it.
Original image availability. Say that unprocessed originals are held and available, as Wiley expects and Elsevier may request.

Worked example (illustrative only)

This wording is invented for illustration. It is not a publisher template, and every name, number and date in it is a placeholder.

Figure 2 (conceptual schematic of the sampling workflow) was drafted with [TOOL NAME, VERSION] on [DATE] using the prompt held in the supplementary file. The output was redrawn and checked against the study protocol by the first and corresponding authors; labels and arrows were corrected by hand. No AI tool was used to create or alter Figures 1, 3 or 4, which show unprocessed or conventionally adjusted original data. The authors confirm that they hold the rights to publish the final figure and take responsibility for its accuracy.

For a ready-made library of wording across publisher families, see AI disclosure statement examples.

Provenance and labelling standards

Two technical standards describe how an image’s origin can be recorded in the file itself. Neither replaces a disclosure statement, and none of the publisher pages checked for this guide state a requirement to attach either one.

C2PA Content Credentials. The Coalition for Content Provenance and Authenticity publishes an open technical standard, branded Content Credentials, for recording the history of a piece of content. Version 2.2 of its specification describes a manifest of assertions about an asset, bound with a digitally signed, tamper-evident claim. The specification describes cryptographic verification of who signed what and whether the data was altered; it does not say the recorded claims are true. A signature shows integrity and signer identity, not accuracy. See CASRAI’s explanation of C2PA Content Credentials.
IPTC Digital Source Type. IPTC maintains a controlled vocabulary for the source of a digital image. It includes trainedAlgorithmicMedia (“Created using Generative AI”), compositeWithTrainedAlgorithmicMedia (“Edited using Generative AI”), compositeSynthetic (“Composite including generative AI elements”) and algorithmicMedia (“Pure algorithmic media”). Authors can use these terms to keep their own image records consistent with the roles in a disclosure statement.

The practical reading for editors: a credential or a source-type label is useful evidence when present, but a missing label proves nothing, because metadata can be stripped when a file is exported or re-saved. The C2PA specification itself discusses assets being separated from their manifests.

How do editors and integrity offices verify an AI image claim?

Verification combines documents from the author with technical checks, and no single input decides the case.

Ask for the underlying record. Original, unprocessed images (Wiley), pre-AI-adjusted versions (Elsevier), the tool name, version and prompt, and the source material (Springer Nature).
Compare with the methods and the data. An image that cannot be reconciled with the described experiment or the raw files is a concern whatever tool produced it.
Check provenance metadata where it exists. A valid C2PA manifest or an IPTC digital source type value is a lead, not proof of truth, and its absence is not evidence of anything.
Use screening tools as one input. A commercial AI image detection service such as AI image detection can be run on a file to produce a score. TruthScan’s site describes a commercial platform operated by Undetectable Inc. offering AI image and document detection through an API, dashboard and Chrome extension, with a free tier, and it publishes no accuracy or false-positive rates (checked 2026-10-30). Third-party tools are named as examples, are not endorsed, and have not been validated by CASRAI. A detector output is probabilistic and cannot establish that an image was or was not AI-generated; treat it as a reason to ask questions, never as evidence on its own.
Follow the journal’s process. Record the request, the author’s answer and the decision, and keep the reviewers’ own use of manuscript files within publisher rules (Taylor and Francis and PLOS both bar uploading unpublished manuscript content to generative AI tools).

For the forensic side (duplication, splicing, selective adjustment), see CASRAI’s guide to detecting image manipulation in figures.

What are the consequences of undisclosed or prohibited use?

The pages read are consistent on the principle, though none sets a single penalty. ICMJE states that failing to disclose AI use may require corrective action and may be construed as misconduct in some circumstances. Wiley treats inappropriate image manipulation as data fabrication or falsification. Springer Nature describes visual content without verifiable sources as opaque and not permitted. Disclosure does not rescue a use the policy prohibits: an AI-generated micrograph disclosed in a caption is still barred at a journal that prohibits it.

What follows is handled through the publisher’s own integrity process: a request for originals and explanations, then a correction or a retraction decision by the editor depending on severity. The exact thresholds are set by each publisher, so cite the publisher’s own notice and process in any case file rather than a general rule.

Checklists

For authors

Classify every image: research data, illustration, or routine adjustment of a captured image.
Read the target journal’s policy page and note its date.
Do not use generative AI to create or alter research-data images unless the journal explicitly permits it and describes the method reproducibly.
Keep unprocessed originals, prompts, tool name and version, and dates.
Disclose in each required place: caption, AI declaration, methods, cover letter.
Confirm you have the rights to publish the output and can verify its accuracy.
Make sure every co-author knows how each figure was produced.

For editors and reviewers

State the journal’s image policy in the author guidelines, with a date.
Ask the same questions of every manuscript: tool, role, human review, originals.
Compare disclosure statements against the figures and the methods.
Treat detector scores, and missing or present metadata, as leads, not findings.
Keep unpublished manuscript files out of generative AI tools.
Escalate suspected manipulation to the publisher’s research integrity process and document each step.

Common mistakes

Disclosing AI writing help but not an AI-made figure.
Assuming that disclosure makes a barred use acceptable.
Treating “enhancement” as harmless when it adds or removes a feature.
Naming “AI” without a tool name and version.
Citing one publisher’s rule as if it applied to every journal.
Reading a detector score as proof, or reading a missing credential as proof of fabrication.
Pasting an unpublished manuscript’s images into an online AI tool during review.

Frequently asked questions

Do journals allow AI-generated images?

It depends on the image and the journal. Elsevier does not permit generative AI to create or alter images in submitted manuscripts, except where it is part of the research design and described in the methods. Springer Nature, Wiley, PLOS and Taylor and Francis allow some uses, such as illustrations or data visualisation, under conditions and with disclosure.

Where should the disclosure go?

Springer Nature asks for an AI Declaration and a caption or legend note; PLOS asks for the figure legend; SAGE asks for the methods or acknowledgements; Elsevier asks for a declaration statement above the references. Follow the journal’s instruction and, when unsure, put it in more than one place.

Is AI-altered image data research misconduct?

Wiley states that inappropriate image manipulation is considered data fabrication or falsification, and ICMJE says nondisclosure may be construed as misconduct in some circumstances. Whether a particular case is misconduct is a determination for the publisher or institution.

Can editors reliably detect AI-generated images?

Not from a single tool. Detection scores are probabilistic, provenance metadata can be missing, and so editors rely on originals, methods and author explanations as well.

Do I need C2PA or IPTC metadata?

None of the publisher pages checked require it. It can support your own record keeping and an editor’s review.

Can an AI tool be an author because it made a figure?

No. ICMJE states that AI tools cannot be listed as authors because they cannot take responsibility for the work, and Taylor and Francis states that AI tools must not be listed as co-authors.

Sources

All checked 2026-10-30. Dates in quotation marks are as shown on the page itself.

Springer Nature, AI use in manuscript preparation: https://www.springernature.com/gp/policies/editorial-policies/ai-manuscript-preparation (no policy date shown)
Elsevier, the use of generative AI and AI-assisted technologies in writing: https://www.elsevier.com/about/policies-and-standards/the-use-of-generative-ai-and-ai-assisted-technologies-in-writing-for-elsevier (“Policy updated: July 2026”)
Wiley, ethics guidelines: https://authors.wiley.com/ethics-guidelines/index.html (“Last updated: July 2026”)
Taylor and Francis, AI policy: https://taylorandfrancis.com/our-policies/ai-policy/ (no policy date shown)
SAGE, ChatGPT and generative AI: https://www.sagepub.com/en-us/nam/chatgpt-and-generative-ai (no policy date shown)
PLOS, ethical publishing practice: https://journals.plos.org/plosone/s/ethical-publishing-practice (no policy date shown)
ICMJE, AI use by authors: https://www.icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html (no section date shown)
C2PA, Content Credentials: https://c2pa.org/ (page last updated 2026-01-08) and C2PA Technical Specification 2.2: https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html
IPTC, Digital Source Type vocabulary: https://cv.iptc.org/newscodes/digitalsourcetype/
TruthScan (named as an example only): https://truthscan.com/

Disclosure and Transparency Standards for AI-Generated Images in Academic Publications