In January 2024, the American Association for the Advancement of Science said the Science family of journals would begin using Proofig, an automated image forensics tool, on accepted manuscripts. The policy shifted image screening from a post-publication catch-up exercise into a routine checkpoint before the version of record goes live.
For authors, that changes a familiar sequence. Conditional acceptance has never been a guarantee, but the weeks between acceptance and publication now include software that can flag a duplicated western blot band, a rotated fluorescence panel, or a recycled microscopy field that belongs to a different figure. The message from the journal is not necessarily an accusation. It is increasingly a standard request for source data.
Why journals are moving image checks before publication
The case for early screening is straightforward. A post-publication correction consumes editor time, invites news coverage, and follows authors for years. A pre-publication query is cheaper and quieter, and it protects the scientific record without requiring a retraction.
Publishers have watched the same pattern repeat often enough to know the cost. A paper appears, a reader posts on PubPeer, an independent expert such as Elisabeth Bik notices duplicated bands, and an institutional investigation begins. Moving the inspection before publication does not remove that process, but it intercepts a meaningful share of cases where the problem is accidental or easily explained.
The first large-journal move
The Science family decision carried weight because it covered multiple titles under one policy. Editors receive a similarity report from Proofig when the software detects duplicated, rotated, scaled, or spliced regions. The tool does not decide the outcome. A journal editor reviews the report, compares it with the authors' raw data if needed, and issues a query.
Other publishers have built parallel systems. EMBO Press pairs human image specialists with software checks, and the Journal of Cell Biology has maintained figure integrity screening as part of its production process for years. The approaches differ in workflow, but they share a principle: the pixel-level check belongs before the final version, not after the press release.
Tools that read pixels rather than prose
Proofig works primarily within a single manuscript, searching for regions that repeat after rotation, resizing, partial erasure, or splicing. ImageTwin adds a comparison against published literature and a database of previously flagged images. Neither product proves misconduct. Both produce a shortlist for a human editor, because the distinction between an honest composition error and manipulation depends on context that software cannot read.
The software examines figures in ways a peer reviewer cannot easily replicate. It checks whether two bands in separate blots are pixel-for-pixel identical. It identifies a cell cluster that appears in another panel after a 180-degree turn. It spots cloned background patches that would be invisible at print resolution. That matters because reviewers read figure legends and assess experimental design; they do not typically align pixels across panels.
What automated screening catches in practice
A 2016 mBio study led by Elisabeth Bik, Arturo Casadevall, and Ferric Fang examined 20,621 biomedical papers and found inappropriate image duplications in 3.8% of them. The authors described that figure as a conservative starting point, since their method could only detect certain patterns. Across a literature that adds hundreds of thousands of papers each year, even a low single-digit rate translates into a large editorial workload.
In day-to-day editorial work, flags often resolve as mislabelled panels, an accidentally reused loading control, or a figure assembled from the wrong batch of high-resolution files. Editors who handle these queries say many cases end with a correction before publication rather than a misconduct finding. That is exactly why the tools have spread: they catch process mistakes that later become public controversies.
False positives and the limits of detection
Automated screening is not a truth machine. JPEG compression, legitimate reuse of a loading control that an author forgets to declare, and long exposures can all trigger a similarity report. Authors may spend days retrieving old files, and the burden falls unevenly on researchers who have moved labs, changed institutions, or never built a reliable raw-data archive. A query is still a request for evidence, not a verdict.
AI-generated and paper-mill images create a harder problem. A fabricated image can be produced without duplicating an existing region, which means within-paper matching may miss it entirely. That is why publishers pair image checks with data-sharing rules, provenance questions, and registry systems. As earlier reporting on AI tools that flag doctored data explained, detection is one layer among several.
What authors should do before submission
The practical response is simple. If a journal will screen your figures, screen them first. Free and paid preflight tools exist, but the product matters less than the archive behind it. Keep unprocessed blots, micrographs, and gel images in labelled folders that can survive a lab move. Describe any contrast or brightness adjustments in the methods and apply them to the whole image. Speak plainly in the figure legend when a loading control is reused.
- Run preflight software on assembled figures before you submit.
- Store raw, unedited images with metadata and a file name that maps to each figure.
- Declare any contrast, brightness, or cropping adjustments in the methods.
- Describe reused loading controls and replicate gels explicitly in the figure legend.
- When an editor asks for source data, send the raw file rather than a polished copy.
Early-career researchers who inherit projects should treat raw data handoff as part of authorship transfer. For faculty members, the question is less about whether an image would survive a check. It is about whether a postdoc could find the original file in one afternoon if a query lands during a holiday week.
Corrections, retractions, and paper mills
The connection between image screening and corrections is real but uneven. Some flagged papers receive a formal erratum; others are retracted when the raw data cannot be produced and the editors cannot verify the figure. Missing files are not proof of fraud, but they leave a committee with little room to act. The COPE paper mill detection guidance and COPE's broader publication ethics resources help editors separate honest error from organised misconduct.
Paper mills have raised the stakes. The Wiley and Hindawi mass retractions showed how quickly a publisher can be overwhelmed when post-publication policing carries the entire burden. Automated image checks are not a solution to that crisis, but they are an inexpensive early filter that reduces the inflow.
For journal teams, the workflow creates a new editorial skill. Someone must read a similarity report, write a fair author query, and distinguish a declared reuse from an undeclared one. That skill set draws on publication ethics, data management, and careful communication that does not presume guilt. Larger publishers have built integrity teams; smaller journals may depend on one production editor who has learned the work on the job.
If you are submitting this month, take one small step before you upload. Run your assembled figures through a preflight check and save every raw image in a folder named by figure and panel. The preprint is yours; the published version belongs to the record. In between sits a set of pixels that now tells more of the story than it used to.
Photo by Markus Winkler on Unsplash