Automated image-forensics software has moved from a niche add-on to a standard part of the editorial pipeline at several major publishers over the past year. Two developments make the shift concrete rather than aspirational: MDPI signed a multi-year deal with Proofig AI in September 2025 to screen submissions across its biomedical journals, and the American Society for Microbiology (ASM) published a peer-reviewed account in mBio of a year-long pilot that made Imagetwin a standard part of its ethics screening workflow. Together they give a rare thing in this space: publicly documented, named adoption with real numbers attached, rather than industry-wide estimates.
What image-integrity screening tools actually check
Tools such as Proofig AI and Imagetwin run submitted figures through automated pattern-matching before or during peer review, flagging candidates for human editorial review rather than making final decisions themselves. Reported detection categories, per both vendors’ own published capability lists, typically include:
- Internal duplication — the same photographic element (a Western blot band, a microscopy field, a gel lane) reused within one figure or across figures in the same manuscript, including after rotation, cropping, or contrast adjustment.
- External duplication and plagiarism — a submitted image matching one already indexed from another publication, including cross-database checks against literature repositories such as PubMed Central.
- Manipulation and splicing — signs that an image has been assembled from multiple sources or selectively altered in ways not disclosed in the methods.
- Undisclosed AI-generated content — synthetic figures, including diffusion-model-generated micrographs or blots, presented as original data.
These tools sit alongside, not in place of, the editorial decision frameworks publishers already use once a concern is raised — COPE’s “Inappropriate image manipulation in a published article” flowchart (version 2, published 20 May 2024) is the widely-referenced example, but it is explicitly a procedural tool for routing a flagged case toward correction, expression of concern, or retraction, not a technical detection method itself. Screening software supplies the flag; COPE-style flowcharts and editorial judgment decide what happens next.
MDPI’s multi-year Proofig AI agreement
MDPI and Proofig AI announced their partnership on 25 September 2025, following what both parties described as a successful pilot. Per Proofig AI’s own announcement and trade coverage from ALPSP and STM Publishing News:
- The rollout began across a set of MDPI’s biomedical journals, with stated plans to extend to additional titles.
- During the pilot, Proofig AI’s system identified duplicated confocal microscopy and histology images, including cases where the duplication had been disguised through cropping or other edits.
- Tim Tait-Jamieson, MDPI’s Head of Publication Ethics, and Sanita Meijere, MDPI’s IT Product Manager, are quoted in the announcement discussing the tool’s fit into MDPI’s existing ethics workflow; Proofig AI CEO Dror Kolodkin-Gal is quoted describing the pilot results.
This is worth flagging in context: MDPI has drawn publicized scrutiny over editorial practices in recent years, including a 2014-2015 Beall’s List listing and delisting, and Clarivate’s 2023 discontinuation of Web of Science coverage for two of its journals over a content-relevance review. A formal, named, multi-year integrity-tooling commitment is a concrete, checkable data point against that backdrop — independent of how any single publisher’s broader editorial reputation is assessed, it is a real procurement decision that can be verified against the vendor’s own announcement and independent trade press.
ASM’s Imagetwin pilot: the numbers behind the rollout
ASM’s account, published in mBio (Chaturvedi, Hibbard, Nelson, Casadevall & Kullas, “ASM incorporates Imagetwin to address image duplication and preserve scientific accuracy”), is unusual in publisher-tooling coverage for actually reporting pilot-level data rather than only announcing adoption. Key figures from the paper:
- ASM integrated Imagetwin into its journals’ editorial workflow in 2023 and ran a one-year pilot.
- 2,627 accepted, eligible manuscripts were screened during the pilot.
- Imagetwin flagged image duplication concerns in 3.9% of those manuscripts.
- Most flagged concerns were unintentional (e.g. accidental reuse of a figure panel) and were resolved through author correction before publication.
- Six manuscripts (0.23% of those screened) had their acceptance revoked because the image issues could not be resolved.
- Screening covers halftone images specifically — microscopy, gel, and immunofluorescence images — applied at the revision or resubmission stage alongside other ethics checks such as text-similarity screening. The paper notes the tool does not evaluate line art, sketches, or graphs.
Since the pilot concluded, Imagetwin screening has become a standing part of ASM’s ethics checks rather than a time-limited trial. The 0.23% revocation rate is a useful reference point: it suggests that even with systematic screening in place, outright fraud is a small fraction of what gets flagged — most hits are correctable errors, which is consistent with what image-integrity investigators such as Elisabeth Bik have long argued about the balance of intent behind duplicated figures.
Why this is happening now
Three converging pressures explain why publishers are formalizing tooling rather than relying on ad hoc reviewer vigilance or post-publication sleuthing:
- Volume. Manual visual comparison does not scale to modern submission volumes, particularly at high-throughput open-access publishers.
- The paper-mill and image-fraud landscape. Automated image-similarity screening became widespread across major publishers roughly 2022-2024, arriving alongside text-forensics tools such as the Problematic Paper Screener and STM Integrity Hub — part of the same broader response to organized manuscript fraud. See CASRAI’s companion piece on paper mills and tortured phrases for that wider context.
- Generative AI. A diffusion-model-generated Western blot has no duplicate source image to match against, which is a materially harder detection problem than catching a copy-pasted lane. Vendors are actively extending detection toward synthetic-image flagging, but this is the least mature part of the toolset industry-wide — see CASRAI’s synthetic image and AI image generation dictionary entries for the underlying concepts.
Limitations to keep in view
Neither case study suggests these tools are a complete solution:
- Both vendors’ tools are reported to be strongest on photographic/halftone images (blots, micrographs, gels) and weaker or non-functional on line art, schematics, and graphs — ASM’s paper states this limitation directly.
- Flags require human editorial adjudication; software output is a triage signal, not an automatic verdict. This mirrors how COPE’s own flowchart is scoped — a procedural decision aid for humans, not a detection method.
- Publicly reported adoption remains concentrated among a relatively small number of named publishers and societies. Broader industry-wide uptake figures are not independently verifiable from public sources at this time, and should not be assumed from these two examples alone.
- Detection of undisclosed AI-generated images is the newest and least battle-tested capability either vendor offers; treat marketing claims in this specific area with more caution than the duplication-detection claims, which have a longer track record.
What this means for other publishers and integrity offices
For research-integrity offices and smaller or society-run journals watching MDPI and ASM’s moves, the practical takeaways are: (1) a pilot-then-scale approach, tested on a defined slice of submissions before workflow-wide rollout, is the pattern both organizations followed; (2) screening tools work best layered into an existing ethics-check stage (alongside text-similarity screening) rather than as a standalone gate; and (3) publishing pilot data, as ASM did in a peer-reviewed venue, is itself a meaningful transparency practice that other adopters could follow to make the field’s actual effectiveness rates less opaque than they currently are.
Frequently asked questions
Are Proofig AI and Imagetwin the same kind of tool?
Both are commercial image-forensics platforms marketed to scholarly publishers for pre-publication screening, with substantially overlapping capability sets (duplication detection, manipulation flags, and growing AI-generated-image detection). They are competing products from separate companies, not the same tool under different names.
Does an image-integrity screening flag mean a paper is fraudulent?
No. ASM’s published pilot data found that of manuscripts flagged for image duplication, the substantial majority of concerns were unintentional and resolved through correction; only a small fraction (0.23% of all manuscripts screened) resulted in a revoked acceptance.
Do these tools check for AI-generated text as well as images?
No — Proofig AI and Imagetwin are specifically image-forensics tools. Text-similarity and AI-text-detection screening (and paper-mill signal detection more broadly) run through separate systems; see CASRAI’s coverage of GenAI disclosure in scholarly authorship for that adjacent landscape.
Is image-integrity screening now standard across scholarly publishing?
Automated image-similarity screening became common practice at several large publishers between roughly 2022 and 2024. MDPI’s and ASM’s 2025 moves are concrete, individually verifiable examples of that trend continuing and formalizing, but a precise industry-wide adoption rate is not something CASRAI can independently verify from public sources, and readers should not extrapolate universal adoption from these two cases alone.
This article synthesizes publicly available vendor announcements, trade press, and a peer-reviewed pilot report. It does not represent an endorsement of any commercial product by CASRAI.
