A postdoc I'll call Dr. S spent two nights rebuilding a western blot figure because a reviewer's assistant flagged a duplicated band. The raw file showed the image was clean; the slide template repeated a lane label, not a result. Multiply Dr. S by the thousands of papers that contain something worse. A 2016 mBio analysis of 20,621 biomedical articles found problematic images in 3.8 percent of them, with at least half showing signs of deliberate manipulation. That study set the base rate, and it is the reason publishers now run new AI detection software before reviewers ever see a manuscript.
The tools do not work like a human eye. They are computer vision systems trained to detect duplicated regions, rotations, flips, splices and copy-move edits within a single figure. Proofig and ImageTwin are two commercial systems now used by journals and publishers; the International Association of Scientific, Technical and Medical Publishers (STM) has begun integrating similar image checks through its cross-publisher Integrity Hub. It takes seconds per figure, not minutes or hours. That speed is what changed. A journal that used to sample images can now screen every one.
Why Publishers Are Now Screening Every Figure
The mBio study's 3.8 percent is not a retraction rate; it is a screening yield. That study, led by Elisabeth Bik and colleagues, examined papers from 40 journals and found duplicated bands, re-used microscopy fields, and rotated or resized sections in roughly one of every 26 papers. You can read the full analysis in mBio. The authors estimated that more than half of the flagged images suggested intentional manipulation, while the rest looked like assembly errors. That distinction matters. Software flags a pattern; it does not know intent.
The result is that editors and publishers have moved from sampling to universal screening. A journal that processed perhaps a few hundred images by eye can now run a full issue through software in an afternoon. That increases the chance a sloppy or fraudulent figure gets caught before publication. It also produces false positives, which is why journals keep human review in the loop. The acceptable false-positive rate depends on the field, the figure type, and the stakes of a retraction. In high-volume cell biology journals, the tolerance for missing a duplicated band is close to zero; in clinical case reports, an over-flag can stall a clean submission.
Inside the New AI Screening Tools
The core methods draw from forensic image analysis and computer vision. A detector splits each figure into tiles, compares tiles for similarity under transformations, and maps regions that match more than expected. Some tools also look at blot lanes, microscopy panels, and flow cytometry charts separately because each format fails in different ways. The commercial tool Proofig advertises checks for duplicated, flipped, and rotated image sections. ImageTwin offers a similar approach and has been trialled by publishers for pre-submission screening. The STM Integrity Hub provides shared infrastructure so publishers can screen manuscripts with common tools before editorial review. That shared layer matters because paper mills submit the same fabricated figure to multiple journals.
One limitation: these systems are better at finding duplicated regions inside a single paper than at locating a figure copied from someone else's publication. They need a reference database to catch cross-paper duplication, and those databases are uneven across disciplines. A western blot from a 2008 paper may not be indexed in a way the software can search quickly. The base rate of detectable duplication is therefore lower than the true rate. That gap is a data-infrastructure problem, not a lack of algorithmic ambition.
The Human Review Still Decides
After software flags a figure, an editor or research integrity officer checks the original files and asks the authors for raw data. The COPE digital image guidelines make one point repeatedly: any alteration that changes the interpretation is unacceptable, while minor brightness or contrast adjustments applied to the whole image are usually fine. A duplicated band in one figure can be an honest cut-and-paste error if the corresponding raw file is intact and the underlying data are genuine. If the raw file is missing, the editor's options narrow, and rejection becomes more likely. Retractions typically follow when journals cannot verify the original data.
I watched a neighboring lab handle this badly. A former research assistant had pasted a band from one experiment into another figure and then left the institution. The principal investigator spent four weeks reconstructing what the assistant had done, and the journal retracted two papers. The PI did the right thing after the fact. What the newer AI tools do is move that confrontation closer to submission, before a paper has time to influence a field.
This is part of a broader shift in submission pipelines. The same infrastructure that now runs image checks is being tuned to spot paper mills, a problem covered in earlier reporting on the Wiley and Hindawi mass retractions. Publishers are combining image forensics with author verification and duplicate submission checks. None of these steps alone stops fabrication; together they raise the cost of getting a fake figure through.
What This Means for Your Lab
Treat image integrity as a pre-submission habit, not an afterthought. The practical steps are small, and they align with data-sharing practices most labs already claim to follow.
- Keep raw, unedited image files in a separate folder from processed figures, and never overwrite the raw file.
- Label every processing step in a text file next to the data, including brightness adjustments and cropping. This is the README equivalent for images.
- Run your own figures through an image-checking tool before submission when the journal or your institution offers one.
- For each figure, prepare the raw data and acquisition settings so you can respond to an image query within 48 hours instead of three weeks.
- If a figure contains reused controls, annotate that clearly in the figure legend and the cover letter.
The least comfortable step is the last one: if you spot a problem in a colleague's figure, say it before the manuscript leaves the lab. The cost of an awkward conversation at submission is a fraction of the cost of a retraction after publication. Dr. S's clean-image scare was sorted out in two days because the raw files were named properly and the lab notebook was legible. The PI next door spent four weeks on the same question because those files did not exist.
Start with the next manuscript your lab will submit, not the one already under review. Download the journal's image guidance, check the figure preparation tool, and make one folder called raw_images with one README. If the journal cannot tell you how image queries are adjudicated, ask.
