03 · What You Need to Know
What Researchers Should Report About AI-Assisted Scientific Images
Why Does AI Image Processing Require Methodological Transparency?
Scientific images may function as research evidence rather than merely visual illustrations. Their interpretation can depend on acquisition conditions, processing decisions, and the relationship between recorded signals and displayed structures.
AI image processing introduces additional considerations because some models estimate image content using learned patterns. These transformations may improve analytical performance, but they may also suppress weak signals, modify boundaries, or introduce features that are not independently established by the original measurements.
Readers therefore need sufficient information to distinguish the observations acquired by the researcher from subsequent computational transformations.
For example, reporting that a microscopy image was "enhanced using AI" does not establish whether the model performed denoising, super-resolution, reconstruction, segmentation, or generative filling. These operations have different methodological implications.
The purpose of disclosure is to make consequential processing decisions understandable and assessable, not merely to acknowledge the presence of an AI tool.
Is AI-Use Disclosure the Same as Reporting a Research Method?
No. The two activities overlap but address different questions.
An AI-use declaration identifies the involvement of AI-assisted technology according to applicable editorial requirements. Methodological reporting explains how the procedure was performed and how it contributed to the research.
AI-use disclosure
Identifies the AI technology used, its purpose, and the nature of its contribution according to applicable reporting requirements.
Methodological documentation
Explains the processing workflow, relevant parameters, validation procedures, and effects on scientific measurements or interpretation.
A manuscript may need both. A general declaration that AI was used cannot substitute for a reproducible description of an image analysis procedure.
Conversely, a detailed Methods section does not necessarily satisfy a journal's separate requirement for an AI-use statement.
What Information Should Be Included in an AI Image Processing Disclosure?
The necessary details depend on the technique and scientific endpoint. A useful reporting framework considers the following elements.
| Reporting Element |
Information to Provide |
Why It Matters |
| Tool identification |
Software, model name, version, and relevant implementation. |
Establishes which technology was used. |
| Purpose |
Whether AI performed denoising, reconstruction, segmentation, enhancement, or another task. |
Clarifies the methodological role. |
| Input data |
Relevant image modality, acquisition characteristics, and preprocessing. |
Defines the information supplied to the model. |
| Processing procedure |
Parameters, thresholds, model configuration, and consequential transformations. |
Supports methodological evaluation and reproducibility. |
| Output status |
Whether results represent processed observations, predictions, or generated estimates. |
Prevents confusion between measurement and synthesis. |
| Validation |
Reference evidence, performance metrics, and checks relevant to the scientific endpoint. |
Supports claims about reliability. |
| Researcher oversight |
Relevant review, correction, and acceptance procedures. |
Documents analytical accountability. |
| Limitations |
Known errors, uncertainty, and restrictions on interpretation. |
Prevents overstatement of the evidence. |
This framework is a practical reporting aid rather than a universally prescribed disclosure template. Researchers should adapt it to the actual method and the requirements of the intended journal.
Where Should AI Image Processing Be Reported in a Manuscript?
Different parts of a research paper serve different reporting purposes.
The Methods section is generally appropriate for explaining AI procedures that form part of data preparation, reconstruction, segmentation, measurement, or analysis.
Figure captions may identify particular images that have undergone consequential processing and clarify whether displayed features are original observations or model-dependent estimates.
Supplementary materials can contain additional validation results, processing settings, or methodological details that would be difficult to accommodate in the main text.
A dedicated AI-use declaration may also be required by the publisher, even when the Methods section already describes the procedure.
| Manuscript Location |
Appropriate Information |
| Methods |
Model, processing workflow, parameters, validation, and measurement procedures. |
| Figure caption |
Image-specific processing, generated elements, and relevant interpretation limits. |
| Results |
Validation outcomes and effects of processing on reported measurements. |
| Supplementary materials |
Detailed configurations, additional comparisons, and supporting validation records. |
| AI-use declaration |
Publisher-required identification and description of AI assistance. |
| Data availability statement |
Availability of original images, derived data, and relevant processing records. |
Not every study needs a separate statement in every location. The reporting structure should reflect the significance of the processing and the applicable journal instructions.
How Should AI Denoising Be Disclosed?
When AI denoising is used, researchers should identify the model and explain how it was applied to the scientific images.
Relevant details may include the noise characteristics, model configuration, processing settings, and whether denoising occurred before quantitative analysis.
Where denoising could influence the scientific endpoint, researchers should describe how signal preservation and measurement effects were evaluated.
For example, a study examining faint fluorescence should explain whether the denoising procedure was assessed for its ability to preserve weak genuine signals.
Merely stating that images were cleaned is inadequate when the transformation may change what the study measures.
The risk that AI denoising alters scientifically meaningful information should inform both validation and reporting.
How Should AI Upscaling or Super-Resolution Be Reported?
Researchers should distinguish enlargement for display from model-based estimation of higher-resolution information.
If an AI super-resolution model generates fine structural details, the report should identify the method and clarify the evidential status of those details.
Relevant documentation may include the original resolution, upscaling factor, model version, processing settings, and reference evidence used for validation.
When upscaled images contribute to measurements, researchers should explain how the method was evaluated for the intended endpoint.
A figure caption should not imply that the imaging instrument directly resolved details that were produced through computational estimation.
The possibility of AI upscaling introducing unsupported image details makes this distinction particularly important.
How Should AI Image Segmentation Be Disclosed?
AI segmentation is an analytical procedure, so reporting should focus on how structures were identified and how segmentation performance was established.
Researchers should identify the model, version, preprocessing, relevant inference settings, and any manual correction procedures.
Where segmentation masks produce numerical results, the report should explain how the masks were converted into measurements.
Validation should address errors relevant to the scientific endpoint, such as missed objects, inaccurate boundaries, or differential performance across experimental conditions.
The reporting requirements for AI-assisted scientific image segmentation therefore extend beyond naming the software.
Readers should be able to understand how the reported counts, areas, or volumes were obtained.
What if AI Generates or Fills Missing Image Content?
Generative filling presents a more consequential distinction because the resulting pixels may not correspond to authentic observations.
If a model reconstructs missing regions, researchers should identify which areas were estimated and explain the methodological basis for the reconstruction.
Where appropriate, masks or annotations can distinguish observed regions from model-generated content.
However, transparent reporting does not automatically authorize the use of generated content in primary scientific images.
Researchers must follow applicable restrictions and should not present synthesized structures as directly observed information.
The question of filling missing scientific image details is therefore both a methodological and an evidential issue.
Should Researchers Disclose AI Processing That Was Used Only for Presentation?
Presentation-only processing may still require disclosure when it materially changes the appearance or interpretation of scientific evidence.
For example, an AI model that makes microscopy images visually smoother may also suppress weak structures. Even if no measurements are taken from the processed image, readers may interpret the figure as evidence of the presence or absence of those structures.
Routine display adjustments and consequential model-based alterations may be treated differently under journal policies.
Researchers should assess what the operation actually changed and whether the publication venue requires it to be reported.
The label "presentation only" does not eliminate the need to preserve evidential fidelity.
What Do Major Publishers Require?
Publisher policies differ, and their requirements may change. Researchers should consult the official policy of their intended journal at the time of submission.
Elsevier's generative AI policies for journals distinguish explanatory illustrations, data visualizations, and primary research images. Permitted uses may require figure-caption disclosures, general declarations, or methodological reporting, depending on the image category.
Springer Nature maintains separate guidance on generative AI and scientific images, including restrictions on certain generated visual content.
Nature Portfolio's image integrity guidance emphasizes faithful representation of original data and imposes requirements concerning image processing and original image availability.
The International Committee of Medical Journal Editors recommends that authors disclose relevant AI-assisted technologies and remain responsible for submitted work.
These sources should not be treated as interchangeable policies. A procedure permitted by one publisher may be restricted by another.
Does Disclosure Make Prohibited AI Image Manipulation Acceptable?
No. Disclosure and scientific acceptability are separate questions.
For example, a researcher cannot legitimately erase an unexpected structure from an experimental image and justify the alteration merely by acknowledging the use of generative editing.
Likewise, identifying a generated microscopy image as AI-created does not make it authentic experimental evidence.
The distinction between legitimate image processing and misleading manipulation depends on the effect of the transformation and its scientific justification.
Disclosure helps readers understand what occurred. It does not correct an invalid method or override an editorial prohibition.
How Detailed Should the Figure Caption Be?
A caption should identify consequential processing that readers need to understand the figure, while detailed procedures can be described in the Methods section.
For example, a caption might state that the displayed image was processed using a specified denoising method and refer readers to the relevant methodological description.
Where a figure combines original observations with model-generated representations, the distinction should be clear.
Captions should avoid vague expressions such as "AI enhanced" when the operation materially affects interpretation.
They should also avoid claiming that the processing preserved all information unless validation supports that statement.
What Does an Adequate Methods Description Look Like?
A defensible Methods description identifies the processing workflow in terms that another qualified researcher can evaluate.
For example, it might explain that fluorescence microscopy images were processed using a specified denoising model before segmentation, describe the settings and validation procedure, and state whether quantitative measurements were derived from original or processed images.
If the model was fine-tuned, researchers may need to describe the training data, annotation procedure, and separation of training and evaluation datasets.
Where processing was performed using a proprietary system, exact reproducibility may be limited. The report should nevertheless provide the relevant available information and acknowledge consequential restrictions.
Methodological transparency does not require disclosing protected information unlawfully, but it does require accurately describing the procedure and its limitations.
Should Researchers Preserve Prompts and AI Processing Settings?
When prompts influence the output of a generative image system, retaining relevant prompts may help document how the image was produced.
For scientific processing models, other parameters may be more important than natural-language prompts. These can include model weights, software versions, thresholds, random seeds, calibration settings, and preprocessing procedures.
Researchers should preserve the information needed to reconstruct consequential transformations where feasible.
However, exact reproducibility may not always be possible, particularly when commercial services change models without preserving access to previous versions.
Such limitations should be considered when selecting tools for research applications.
How Should Researchers Document Original and Processed Images?
A reliable provenance record connects each processed image to the authentic source data.
For example, a published figure panel should be traceable to the original acquisition file, the processing procedure, and any relevant intermediate outputs.
Researchers should generally preserve original scientific images and the metadata needed to interpret them.
Where AI segmentation or enhancement contributes to quantitative results, retaining the relevant masks, measurements, and validation records can support subsequent verification.
Documentation should comply with applicable data retention, privacy, and research ethics requirements.
Can Researchers Use a Generic AI Disclosure Statement?
A generic statement may be insufficient when AI processing forms part of the research methodology.
For example, "The authors used artificial intelligence to assist with image analysis" does not identify the model, analytical task, or validation procedure.
It may also leave readers uncertain whether AI altered the scientific evidence or merely assisted with figure preparation.
Researchers should distinguish general AI assistance from methods that materially affect the reported results.
Where required, a concise declaration can accompany a detailed Methods description rather than attempting to place every technical detail in one statement.
What if AI Processing Was Performed by a Collaborator or External Service?
Researchers remain responsible for understanding and accurately reporting methods that contribute to their findings.
If a collaborator or service provider processes scientific images, the research team should obtain sufficient information about the software, model, settings, and validation procedures.
Black-box processing may create difficulties when the provider cannot explain how the images were transformed.
Researchers should consider whether the available documentation is adequate for the scientific purpose before relying on the processed outputs.
Outsourcing the processing does not transfer responsibility for the published interpretation to the service provider.
How Can Researchers Verify That Their Disclosure Is Accurate?
Before submission, compare the disclosure with the actual processing records.
Confirm that the named model and version were used, that the stated purpose matches the procedure, and that any claims about validation are supported by documented results.
Check whether the caption, Methods section, AI-use declaration, and data availability statement are consistent.
For example, a caption should not describe an image as unprocessed if the Methods section indicates that it underwent AI denoising.
Researchers should also verify that the reported processing complies with the journal's current policy.
Watch Out
A disclosure statement should not claim that AI processing preserved all scientific information unless that claim has been appropriately validated. Transparency about the tool is not evidence that the transformation was accurate or permissible.
