03 · What You Need to Know
How AI-Assisted Scientific Image Integrity Screening Works
What Is Scientific Image Falsification?
Scientific image falsification occurs when research images or their associated presentation are manipulated in ways that misrepresent the underlying observations or research process.
Under the United States Public Health Service research misconduct framework, falsification includes manipulating research materials, equipment, or processes, or changing or omitting data or results, such that the research is not accurately represented in the research record.
Image-related concerns may include selectively removing experimental features, presenting duplicated images as independent observations, or altering image content in ways that change the scientific interpretation.
However, not every image adjustment constitutes falsification.
Scientific imaging often requires legitimate processing, including adjustments to brightness, contrast, color mapping, cropping, or background correction. Whether a particular adjustment is acceptable depends on the imaging method, its effect on the represented evidence, and applicable disciplinary or journal requirements.
The relevant question is whether the image remains an accurate representation of the underlying data and whether material processing has been disclosed appropriately.
Image Processing, Image Error, and Falsification Must Be Distinguished
Researchers should separate three different situations that may appear similar during automated screening.
Legitimate image processing
Documented adjustments that preserve the relevant scientific information and comply with applicable methodological and publication requirements.
Image preparation error
An incorrect crop, mislabeled panel, accidental duplication, or other mistake that may misrepresent information without necessarily constituting misconduct.
Potential falsification
An alteration or omission that materially misrepresents research evidence and may satisfy the applicable misconduct definition when the required evidentiary and culpability elements are established.
These distinctions matter because a visually suspicious feature does not reveal how it arose.
A duplicated panel may result from an accidental figure assembly mistake. It may also reflect deliberate substitution of an image to support a claimed result.
Determining which explanation is supported requires evidence beyond the visual similarity itself.
What Image Features Can AI Help Detect?
AI-assisted screening may use computer vision, pattern matching, feature extraction, or machine-learning methods to identify suspicious image relationships.
Different methods are designed for different problems, and their performance should be evaluated separately.
| Image feature |
Possible detection approach |
Important limitation |
| Duplicated regions within an image |
Patch matching or copy-move detection |
Repeated biological structures may resemble duplicated regions |
| Similar panels across figures |
Image similarity search and feature matching |
Legitimate reuse or shared controls may require contextual interpretation |
| Rotated or resized duplicates |
Transformation-tolerant feature matching |
Resampling and compression can affect detection |
| Possible local alterations |
Image forensic analysis of inconsistencies |
Acquisition and processing artifacts can create false signals |
| Unexpected image-background patterns |
Texture or statistical anomaly analysis |
Background correction and instrument characteristics may explain patterns |
| Suspicious figure assembly |
Panel segmentation and cross-panel comparison |
Scientific interpretation requires knowledge of experimental conditions |
Detection systems may combine several approaches, but no single method is suitable for every scientific image type.
An algorithm evaluated on ordinary photographs should not automatically be assumed reliable for microscopy, electrophoresis, radiology, or other specialized scientific images.
Duplicated Image Regions Are Important but Not Definitive Evidence
One of the most recognizable image integrity concerns involves duplicated regions presented as though they represent different observations.
For example, two microscopy panels may contain identical cellular arrangements despite being labeled as different experimental conditions.
Such similarity can warrant investigation, particularly when the panels are supposed to represent independent samples.
However, some studies legitimately reuse reference images, controls, or representative panels when the reuse is scientifically appropriate and transparently disclosed.
Researchers should examine the figure legends, experimental design, original image files, and publication requirements before interpreting duplication.
The question is not merely whether two regions match. It is whether their presentation accurately represents the underlying experiment.
Western Blot Images Present Particular Verification Challenges
Western blots and related gel-based images are frequently examined during research integrity screening because bands, lanes, backgrounds, and panel arrangements may contain evidence of inappropriate manipulation.
Potential concerns include duplicated bands, unexplained splicing, selective removal of lanes, or presentation of the same experimental image as evidence for different conditions.
Nevertheless, legitimate figure preparation may involve cropping or combining images when the procedure is scientifically appropriate and clearly documented.
Some journals require disclosure of splicing and access to uncropped original images.
AI may identify visual similarities or discontinuities, but it cannot determine the experimental legitimacy of every arrangement from the published figure alone.
Reviewers may need original exposures, acquisition records, and information about how the final figure was assembled.
Microscopy Images Can Contain Naturally Repeating Patterns
Microscopy images may contain structures that resemble one another because of the underlying biology, imaging technique, or sample preparation.
Cells, tissue arrangements, particles, and other features can produce repeated textures or shapes.
An AI system trained to identify duplicated regions may mistakenly flag these naturally occurring similarities.
Conversely, genuine duplicated regions may be difficult to identify when they have been resized, rotated, cropped, or subjected to additional processing.
Researchers should evaluate whether the detector has been tested on images comparable to those being screened.
Image type, resolution, compression, acquisition settings, and processing history can materially affect the reliability of automated analysis.
Compression and File Conversion Can Complicate Forensic Analysis
Scientific images may pass through multiple software applications before appearing in a manuscript.
Files may be resized, compressed, converted between formats, or incorporated into composite figures.
These operations can alter pixel values and metadata, potentially affecting forensic indicators.
For example, JPEG compression introduces artifacts that may complicate the interpretation of local image inconsistencies.
Researchers should therefore distinguish evidence contained in the original acquisition file from artifacts introduced during subsequent preparation.
Where possible, image integrity assessment should include the original or minimally processed data rather than relying exclusively on a compressed manuscript figure.
Original Image Files Are Central to Verification
Original image files can provide information that is unavailable in the final published figure.
Depending on the imaging system, they may contain acquisition metadata, timestamps, instrument settings, image dimensions, and other contextual information.
Such records can help reviewers reconstruct how a figure was created and determine whether an apparent discrepancy reflects legitimate processing or misrepresentation.
However, metadata alone do not prove authenticity. Metadata may be incomplete, altered during conversion, or unavailable because of the instrument or workflow.
Researchers should evaluate original files alongside laboratory records, processing scripts, figure legends, and other relevant documentation.
Maintaining appropriate image provenance is therefore an important preventive measure, not merely a response to allegations.
Can AI Detect Image Manipulation That Is Invisible to Human Reviewers?
Automated image analysis may identify similarities or numerical patterns that are difficult to notice through ordinary visual inspection.
For example, feature-matching algorithms may identify corresponding regions despite changes in scale or orientation.
This capability can make AI-assisted screening valuable, particularly when editors must examine large numbers of figures.
However, detecting a subtle similarity does not establish its scientific significance.
A detector may identify a technically correct match that is irrelevant to the research claim, or it may miss an alteration that materially changes the evidence.
Researchers should assess both detection performance and the consequences of acting on a flag.
Human expertise remains necessary to determine what the flagged feature means within the experiment.
False Positives Can Harm Researchers and the Scientific Record
A false-positive image integrity flag occurs when a system identifies a suspicious feature even though the targeted problem is absent under the applicable reference standard.
False positives may arise from repeated structures, image-processing artifacts, legitimate reuse, or limitations of the detection algorithm.
When screening results are treated as accusations, the consequences may include publication delays, reputational harm, and unnecessary investigations.
Researchers should distinguish a possible image anomaly from a verified discrepancy and from a formal finding of misconduct.
The risk of AI-generated false accusations of research misconduct requires careful communication and appropriate procedural safeguards.
Editors should avoid public allegations based solely on automated image flags.
False Negatives Can Create Unwarranted Confidence
An image may pass automated screening despite containing a serious integrity problem.
Detectors may fail when alterations are outside their training or detection capabilities, when image quality is poor, or when the relevant evidence is not visible in the submitted figure.
For example, an image may accurately represent its pixels while being assigned an incorrect experimental label.
Such a problem may not be detectable through pixel-level analysis alone.
A negative AI screening result therefore does not certify that the image is authentic, correctly labeled, or scientifically valid.
Researchers and editors should not treat image screening as a substitute for methodological review or appropriate data provenance.
Image Anomalies Cannot Establish Researcher Intent
Formal research misconduct findings may require evidence concerning how the image was altered and whether the relevant culpability requirements have been met.
Under the applicable United States Public Health Service framework, misconduct findings require specified evidentiary elements, including conduct committed intentionally, knowingly, or recklessly.
An AI detector generally cannot determine a researcher's state of mind from the image alone.
Even a verified inappropriate alteration may require further investigation to establish responsibility and whether it constitutes misconduct under the applicable framework.
Researchers should therefore distinguish technical image-forensic findings from conclusions about individual conduct.
The broader distinction between AI-based suspicion and evidence of research misconduct remains essential.
How Should an AI Image Integrity Flag Be Investigated?
A proportionate investigation should begin by identifying the exact feature flagged by the system.
Reviewers should determine whether the concern can be reproduced using appropriate image analysis methods and whether the detector's assumptions apply to the material.
They should then examine relevant source evidence, including original files, figure legends, experimental records, and documented processing procedures.
Where appropriate, authors may be asked to provide clarification or original images under the journal's established procedures.
The response should distinguish correctable figure preparation mistakes from concerns that warrant formal research integrity assessment.
The appropriate human review after an AI research-integrity flag depends on the seriousness and evidentiary strength of the concern.
Should Journals Use AI Image Screening Before Peer Review?
Automated screening may help editors identify issues that deserve examination before or during peer review.
However, a journal should establish how the tool is validated, which image types it supports, and how its findings will be reviewed.
Editors should avoid treating a detector's score as an automatic rejection criterion for misconduct.
Screening may also raise confidentiality and data-processing concerns when unpublished manuscripts are transferred to external services.
Journal policies should address access, retention, permitted use, and the handling of potentially sensitive unpublished material.
COPE guidance can support editorial procedures for addressing concerns, although individual journals and institutions may have additional requirements.
Preventive Image Integrity Practices Are Preferable to Retrospective Detection Alone
Researchers can reduce image integrity problems by preserving original acquisition files, documenting processing steps, and maintaining clear links between images and experimental records.
Figure preparation procedures should be consistent with applicable disciplinary and journal standards.
Researchers should avoid selective alterations that change the scientific meaning of the image and should disclose material processing where required.
Independent checks before submission may identify mislabeled panels, accidental duplication, or inappropriate processing.
AI may assist with these checks, but it should not replace source-data verification.
Watch Out
Do not interpret a duplicated region, suspicious background, or image-forensic score as proof of falsification. Verify the feature against original evidence and consider legitimate processing, figure preparation mistakes, and the applicable misconduct standard.
