AI-generated scientific images are entering the published literature, and visual inspection alone cannot reliably catch them. Proofig AI’s AI-Generated Image Identification capability is built specifically to surface suspected synthetic images across scientific image types for human review before submission.
Why AI-Generated Scientific Images Are a Growing Integrity Risk
Generative AI tools can now produce convincing scientific figures that were never captured in a laboratory. Nature has reported on the rise of AI-generated images in scientific manuscripts, noting that these tools can create figures difficult to distinguish from real microscopy and other modalities. The issue is increasingly recognized across the scientific community and requires a coordinated response from researchers, publishers, and institutions.
The problem extends beyond hypothetical risk. Multiple published papers containing AI-generated figures have been documented, including in high-profile journals, and detection remains a significant challenge. These synthetic images passed peer review, were published, and entered the scientific record before anyone flagged them.
This matters because image integrity problems already affect a substantial share of the life-science literature. By one estimate from an image integrity analyst, about 20 to 35% of submitted manuscripts are flagged for image-related issues during editorial screening. AI-generated images add an entirely new category of risk on top of these existing problems.
Why Visual Inspection Is Not Enough
Modern generative models produce images that mimic the noise patterns, color distributions, and structural features of real scientific data. This is not a failure of reviewers. It reflects how sophisticated these tools have become.
A peer-reviewed study at the University of Split tested medical students and researchers on their ability to detect image duplications, a more established integrity problem than AI generation. Students correctly identified only 29.4% of duplications, and researchers managed just 32.4%. If trained scientists struggle to spot duplications, where at least two copies of the same region exist for comparison, detecting a wholly fabricated image with no original counterpart is even harder.
The generative AI landscape also changes rapidly. New models appear regularly, each producing outputs with different visual signatures. A detection approach that works for one model may miss images from the next. This is why any viable detection system must be continuously updated, not a static filter applied once.
Which Scientific Image Types Are Most at Risk?
AI-generated image risk is not uniform across all figure types. The following categories are both technically feasible targets for generative models and consequential for research claims:
| Image Type | Why It Matters | AI Generation Risk |
|---|---|---|
| Microscopy (confocal, light, electron) | Underpins cell biology, neuroscience, and materials science findings | High: generative models can replicate complex cellular structures |
| Western blots and gels | Central to protein expression and molecular biology claims | High: band patterns are relatively structured and reproducible by AI |
| Histology | Foundation of pathology and tissue-level analysis | High: tissue architecture can be synthesized convincingly |
| Cell plates | Used in drug screening, viability assays, and phenotypic studies | Moderate to high: colony patterns and staining are learnable |
| Medical scans | Support clinical and translational research conclusions | Moderate: imaging modalities have characteristic noise profiles AI can mimic |
| Animal imaging | Documents in-vivo experimental outcomes | Moderate: less standardized, but still vulnerable |
Proofig AI’s detection capability spans all of these categories. This breadth matters because a manuscript often contains multiple figure types, and a check limited to one modality leaves gaps.
How Proofig AI Detects AI-Generated Scientific Images
Proofig AI’s AI-Generated Image Identification is designed to flag suspected synthetic images across the scientific image types listed above. The system identifies images created by the most widely used generative models and is continuously updated as new models emerge. This continuous-update approach is essential given how quickly the generative AI landscape shifts.
Detection results are framed as suspected AI-generated images surfaced for human review. Proofig does not make final determinations about whether an image is synthetic. The goal is to support human assessment by highlighting figures that warrant closer inspection, not to assign intent or guilt.
Proofig’s detection capabilities have been validated on large datasets of real scientific images, and the broader image integrity checks are tuned to produce very few false alarms so researchers reviewing flagged items are not buried in noise.
The credibility of this approach is reflected in its adoption. The Science family of journals adopted Proofig across all six journals to detect image integrity issues before publication, after a multi-month pilot demonstrated effectiveness. MDPI also entered a multi-year partnership, with Milos Cuculovic, Head of Technology Innovation at MDPI, noting: “From the experts, it seemed Proofig AI was by far the best choice, and here we are: further ensuring research integrity and guaranteeing the high quality of research output published by MDPI.”
How AI-Generated Image Detection Differs from Other Image Integrity Checks
AI-generated image detection is a distinct problem from duplication detection, manipulation detection, and image plagiarism checking. Each addresses a different failure mode:
- Duplication and reuse analysis identifies when the same image region appears more than once within a manuscript, whether through scaling, rotation, flipping, or partial overlap.
- Manipulation detection identifies alterations within a single image, such as cloning, splicing, or content-aware edits.
- Image plagiarism identifies reuse of images from previously published manuscripts.
- AI-generated image detection identifies figures that were never captured from real experimental data at all.
A manuscript can pass duplication and manipulation checks and still contain a wholly fabricated AI-generated figure. This is why researchers benefit from comprehensive image integrity quality assurance that covers all of these categories, not just one.
Why Researchers Should Screen Before Submission
The risk of AI-generated images entering a manuscript is not limited to deliberate misconduct. A researcher who unknowingly includes a synthetic figure contributed by a collaborator, student, or co-author faces the same consequences: rejection, correction, or retraction. Even researchers who uphold the highest ethical standards may unknowingly submit problematic images because of how data moves through complex, multi-contributor workflows.
Nature has reported on the personal and career stress researchers face when retracting papers, often due to issues that pre-submission screening could have caught. Retractions carry significant financial and reputational consequences for individuals and institutions alike. More than 8,000 highly cited scientists have at least one retracted publication. Retraction risk spans all tiers of the research community, underscoring why pre-submission screening is a practical safeguard regardless of career stage.
Pre-submission screening gives researchers the opportunity to catch and address suspected issues before they become public problems.
What Researchers Should Do Now
Running a pre-submission image integrity check that includes AI-generated image detection is a practical step any researcher can take today. The workflow is straightforward:
- Upload your manuscript (PDF or image files) to Proofig AI.
- Review flagged items, including suspected AI-generated images, duplications, and manipulations, using the forensic tools provided.
- Export a report for your records or to include with your submission.
The goal is not to assume wrongdoing. It is to protect your work, your name, and the integrity of the research record by surfacing suspected issues early, when they can still be investigated and resolved. Human review of any flagged items remains essential. Proofig surfaces what warrants a closer look. The final judgment belongs to the researcher and, ultimately, to the editorial and peer review process.
