03 · What You Need to Know
How Generative AI Can Compromise the Integrity of Scientific Images
Why are scientific images different from ordinary illustrations?
Scientific images often function as records of observations rather than merely as visual communication materials.
A microscopy image may document cellular structures. An electrophoresis gel may provide evidence concerning molecular material. A medical image may support observations about anatomical or pathological features.
When these images are presented as empirical evidence, readers reasonably expect them to correspond to the underlying observations and acquisition procedures.
Scientific image integrity therefore concerns more than visual quality. It concerns whether the displayed information accurately represents the data obtained.
An image can look professionally prepared while being scientifically misleading. Conversely, an imperfect image may remain valid evidence when its limitations are properly documented.
What constitutes scientific image falsification?
The United States Office of Research Integrity (ORI) defines falsification as manipulating research materials, equipment, or processes, or changing or omitting data or results so that the research is not accurately represented in the research record.
For scientific images, this may involve altering the visual record in ways that create a misleading representation of the observations.
Examples include removing relevant experimental features, introducing structures that were not observed, or selectively modifying image regions to support a preferred conclusion.
The misconduct concern is not determined solely by which software was used. A misleading alteration may be performed manually or through an AI system.
Formal findings of research misconduct also require the applicable standards concerning evidence, culpability, and institutional procedures to be satisfied.
How can generative AI alter the meaning of an experimental image?
Generative AI can produce image content based on learned patterns rather than exclusively preserving recorded pixel values.
Depending on the system, an AI-assisted operation may generate plausible structures, suppress features, reconstruct missing regions, or transform the appearance of an image.
These capabilities can be useful for certain visualization or methodological purposes, but they also create risks when generated content is presented as directly observed evidence.
| AI-assisted operation |
Potential scientific concern |
What researchers should establish |
| Generative image enhancement |
Structures may be introduced or modified during processing. |
Whether scientifically relevant features remain faithful to the source data. |
| Object removal |
Relevant experimental features may disappear. |
Whether removed content affects the evidentiary interpretation. |
| Image reconstruction |
Generated details may be mistaken for observed details. |
Which features are measured, inferred, or reconstructed. |
| Synthetic scientific image generation |
Entirely artificial images may resemble genuine observations. |
Whether the images are clearly identified as synthetic. |
| AI-assisted segmentation |
Incorrect boundaries may affect measurements. |
Whether segmentation accuracy has been validated for the intended application. |
| AI-based denoising |
Weak signals or important details may be suppressed or altered. |
Whether processing changes scientifically meaningful information. |
The existence of these risks does not mean every AI-based operation produces false information. The outcome depends on the method, parameters, source data, and validation.
Is AI-assisted image enhancement always falsification?
No. Scientific image processing has long included legitimate adjustments and computational methods.
For example, researchers may apply appropriate brightness or contrast adjustments, background correction, denoising, or segmentation to support analysis and presentation.
However, such operations should follow the relevant discipline's standards and the requirements of the journal or publisher.
Nature Portfolio's image-integrity guidance addresses inappropriate image manipulation and emphasizes the importance of accurate representation and access to original data.
Adjustments that selectively obscure, enhance, or remove scientifically relevant information may be unacceptable even when the resulting image appears clearer.
AI introduces an additional complication because some processing methods can generate plausible content that was not directly captured by the instrument.
Researchers should not assume that an enhanced image remains evidentially equivalent to the original simply because the software describes the operation as improvement.
What is the difference between denoising and inventing image features?
Denoising attempts to reduce unwanted variation or interference in recorded images. Depending on the method, it may improve visibility or support quantitative analysis.
Generative reconstruction can go further by estimating or producing image features based on patterns learned from other data.
These operations are not necessarily inappropriate, but their evidentiary implications differ.
Validated image processing
A documented procedure that transforms recorded data for a defined scientific purpose, with its effects and limitations appropriately evaluated.
Misleading generative alteration
An operation that introduces, removes, or changes scientifically relevant features and presents the result as though those features were directly observed.
A reconstructed feature may be useful as a model-based estimate. It should not automatically be treated as an independent experimental observation.
When image processing affects quantitative measurements or diagnostic interpretations, validation becomes particularly important.
Can AI-generated microscopy images be mistaken for genuine observations?
Yes. Generative models can produce images resembling microscopy and other scientific imaging modalities.
Such images may be useful for software development, methodological testing, or educational demonstrations when clearly identified as artificial.
However, an AI-generated microscopy image does not establish that a particular cell, tissue structure, or experimental condition was observed.
Presenting a synthetic image as an authentic experimental micrograph would misrepresent its provenance.
This is closely related to AI-assisted research data fabrication, particularly when an entire image is invented rather than an existing observation being altered.
The distinction between fabrication and falsification may depend on whether the research record contains wholly invented evidence or manipulated genuine evidence.
Can AI-assisted image editing affect quantitative research results?
Yes. Scientific images are frequently used to derive numerical measurements, including object counts, fluorescence intensity, tissue area, particle dimensions, and other quantitative indicators.
If AI processing changes the visual features from which those measurements are obtained, the resulting numerical values may also change.
For example, an AI-based segmentation method may incorrectly identify boundaries, leading to inaccurate area measurements.
Such errors do not automatically constitute misconduct. They may reflect limitations of the analytical method.
However, knowingly presenting altered measurements as accurate, concealing consequential transformations, or manipulating results to support a preferred conclusion may raise falsification concerns.
Researchers should validate image-analysis methods and preserve sufficient information to reconstruct the analytical process.
What do journal and publisher policies say about image manipulation?
Policies vary, but major publishers commonly distinguish acceptable image processing from alterations that misrepresent scientific evidence.
Nature Portfolio provides image-integrity guidance concerning appropriate adjustments, image preparation, and original data.
Journal of Cell Biology has also published influential guidance on digital image manipulation, including the need to avoid selective alterations that change the interpretation of experimental results.
Researchers should consult the specific journal's current instructions before preparing figures.
Some publishers impose additional restrictions on generative AI-created or AI-altered images. Requirements may differ for graphical abstracts, conceptual illustrations, and images presented as research data.
Permission to use AI-generated artwork does not necessarily establish permission to use generative AI to alter experimental evidence.
Can AI-generated scientific illustrations be legitimate?
Yes. A conceptual diagram illustrating a proposed mechanism serves a different purpose from a microscopy image documenting an observed structure.
Researchers may use AI-assisted illustration tools to communicate scientific concepts when permitted by applicable policies and when the result is accurately labeled.
However, conceptual illustrations should not be presented as empirical observations.
A generated drawing of a cellular mechanism, for example, may help readers understand a hypothesis. It does not demonstrate that the mechanism was observed experimentally.
The distinction should remain clear in figure captions and the surrounding manuscript.
How can researchers preserve scientific image provenance?
Image provenance concerns the origin of an image and the sequence of processing steps applied to it.
Researchers should preserve original acquisition files wherever possible, including relevant metadata and instrument information.
Processed images should be traceable to their source data through documented procedures.
For AI-assisted processing, useful records may include the software or model used, relevant version information, processing settings, input files, and validation procedures.
When generative reconstruction is involved, the manuscript should distinguish directly observed information from inferred or generated content.
These practices support reproducibility and facilitate investigation of discrepancies without assuming that every unusual image indicates misconduct.
Can AI-assisted image falsification be detected reliably?
Not in every case. Image-forensic methods may identify duplicated regions, inconsistent patterns, or other suspicious features, but no single method establishes the authenticity of every scientific image.
AI-generated alterations may also complicate visual inspection because generated structures can appear plausible.
However, apparent irregularities are not proof of misconduct. Legitimate image processing, acquisition artifacts, and file-handling procedures can produce unusual visual characteristics.
Comparison with original acquisition files and processing records is often more informative than examining a final figure alone.
The broader limitations of detecting AI-assisted research falsification reinforce the importance of preserving original evidence.
Who is accountable when AI changes scientific evidence?
Human researchers remain responsible for the scientific claims they make using AI-processed images.
An automated image-processing system does not assume responsibility for the accuracy of a published figure.
The International Committee of Medical Journal Editors emphasizes that human authors remain accountable for the accuracy and integrity of material produced with AI assistance.
Nevertheless, an individual misconduct finding requires examination of the person's actions and the applicable evidentiary standards.
The principle of human accountability when AI performs a prohibited research action applies even when the software completes the immediate manipulation.
