A BMJ study published this year ran machine-learning screening across millions of cancer research papers and found paper mill signatures at substantial scale. Not a few bad actors. Not isolated misconduct. A detectable, recurring pattern across a corpus that shapes how oncologists make decisions. That result should stop you cold.
Peer review was designed for a world where fraud was rare and expensive to commit. One researcher, one fabricated dataset, one retraction. The system could absorb that. What it cannot absorb is an organized industry that templates papers, manufactures synthetic references, and cycles authorship through revision to dodge detection. Richardson et al. in PNAS 2025 described these entities as "large, resilient, and growing rapidly." That is not a description of misconduct. That is a description of a competing production system.
The Verification Gap Is an Engineering Flaw, Not a Character Problem
Here is the specific failure: peer reviewers evaluate prose quality before they evaluate evidence. The Manusights Research Brief from March 2026 put it plainly: "Evidence is often judged after prose quality, not before." Paper mills exploit exactly that gap. Their output reads well. It cites real journals. It follows template structures that mimic legitimate methodology sections. A human reviewer working on a 2-week turnaround with no access to full-text reference deposits has almost no chance of catching it.
I will grant the skeptics one fair point: not every retraction signals a mill. Honest errors exist. Redundant publication happens without malice. But the BMJ result, combined with the Elsevier energy journal retracting 6 papers in March 2026 for unauthorized authorship changes during revision, and 50-plus flagged papers on superheavy elements, suggests the volume of organized fraud has crossed a threshold where individual reviewer judgment is simply the wrong tool for the job.
Think about what this means downstream. A business school paper cited 2,000 times per year with false claims shapes policy. Cancer research with fabricated data shapes treatment protocols. The cost is not abstract. It lands on patients and on the researchers doing legitimate work who now compete for credibility in a polluted literature.
The Fix Already Exists. Deploy It.
ML screening at the corpus level works. The BMJ study proved it. Full-text reference deposits, which let automated tools verify citations before publication rather than after, are technically straightforward. Pre-publication authorship verification against institutional records is not a moonshot. These are engineering solutions to an engineering problem, and the tools are ready.
The tension I keep running into: deploying these systems costs money and slows submission pipelines, and journals compete for submissions. There is a real incentive to keep friction low. That incentive is exactly why individual journals will not fix this voluntarily. Publishers like Elsevier and Springer Nature need to mandate ML screening across their portfolios, not as a pilot program, not as an optional service tier. Mandatory. Pre-publication. Every submission.
Funding bodies should require it too. If your grant-funded research lands in a journal without verified screening, that is a compliance gap, not a publishing preference.
Peer review is not broken because reviewers are lazy or corrupt. It is broken because it was never designed to detect industrial-scale forgery. The BMJ's algorithm found what 10,000 reviewers could not. Use the algorithm.
