A Publisher Turns AI Against AI at Industrial Scale
On June 18, the open-access publisher MDPI said it has begun running AI-driven research-integrity screening across its full journal portfolio, scanning roughly 2,000 manuscripts every day. The system, called Ethicality, had previously been limited to selected titles; the expansion makes it one of the largest production deployments of automated integrity tooling in scholarly publishing. The move is a direct response to a problem that scaled faster than any human review process could absorb: generative AI has made it trivial to produce plausible fake papers, fabricated references, and laundered text, and the volume now overwhelms manual checks. MDPI's answer is to fight machine-generated fraud with machine-speed detection.
We see this as a preview of what every high-volume content pipeline will eventually have to build. Scholarly publishing is simply the early casualty, because its incentives reward volume and its trust model depends on a fragile assumption that submissions are made in good faith. Dr Enric Sayas, the product owner for Ethicality, did not soften the framing: "We are in a technological race. As generative AI makes it easier to produce sophisticated plagiarism and high-quality fake papers, traditional detection methods are no longer sufficient." That sentence applies just as well to enterprise knowledge bases, marketplace listings, and any system that ingests human-submitted documents at scale.
What Ethicality Actually Inspects
Ethicality does not read a paper the way a reviewer does; it looks for statistical and structural patterns that human eyes miss. The system examines titles, abstracts, author metadata, the main text, and reference lists, then monitors a submission continuously from initial upload through the final publication decision. It is tuned to surface suspected paper-mill activity, fabricated submissions, AI-generated or manipulated writing, fake references, citation irregularities, unusual authorship patterns, and suspicious peer-review behavior, including AI-generated review text. In other words, it watches both the manuscript and the people and processes around it, treating integrity as a property of the whole workflow rather than a single document.
Crucially, Ethicality runs alongside two established third-party tools rather than replacing them. Proofig handles image-manipulation detection, the domain where doctored Western blots and duplicated microscopy images have produced some of the most damaging retraction scandals of the past decade. iThenticate continues to catch duplicated text and conventional plagiarism. Layering a pattern-detection engine on top of these point solutions reflects a hard-won lesson: no single detector is sufficient, and a defensible integrity stack now needs overlapping methods that catch different failure modes. For any organization building content-verification, this composability is the design pattern to copy.
Humans Still Make the Call
MDPI is careful to position Ethicality as a triage system, not a judge. The tool generates risk signals that route flagged submissions to editors or dedicated research-integrity specialists, who decide whether an investigation or any action is warranted. The presence of AI-generated text or an unusual pattern does not, by itself, establish misconduct. Dr Milos Cuculovic, MDPI's head of technology innovation, drew the line explicitly: "AI, when used responsibly, acts as a set of guardrails rather than a substitute for human judgment." That distinction is not just diplomatic cover. It is the only legally and ethically defensible posture when the cost of a false accusation against a researcher is career-altering.
The practical value of the automation is in reclaiming human attention. By offloading repetitive technical work, reference validation, formatting checks, and initial triage, the system frees editors to spend their limited time on questions of genuine research quality. This is the same value proposition that AI vendors pitch across every knowledge-work function, but here it carries unusual weight, because the alternative is editorial teams drowning. As Cuculovic put it, "traditional, manual processes are no longer sufficient in peer review." The lesson for enterprise leaders is that AI integrity tooling earns its keep not by replacing judgment but by protecting the scarce capacity to exercise it.
The Stakes for the Scientific Record
The reason this matters beyond publishing is that the scientific record is upstream of almost everything else. Corporate research, regulatory decisions, clinical guidance, and increasingly the training data for the next generation of AI models all draw on published literature. If paper mills and synthetic studies contaminate that corpus at scale, the pollution propagates into every downstream system that trusts it, including the models that will be trained to detect the very fraud they helped enable. MDPI, which has itself faced criticism over publication volume and quality control, has a clear commercial incentive to demonstrate rigorous gatekeeping, and a 2,000-per-day screening throughput is a credible signal of seriousness.
There is a deeper, recursive irony that technology leaders should sit with. The same generative models that make fraud cheap are now the only economical way to detect it, which means integrity becomes a permanent arms race rather than a solved problem. Detection thresholds will need constant retuning as generation improves, and adversaries will probe for blind spots. Any enterprise that ingests external content, from insurance claims to product reviews to job applications, is now in the same race MDPI describes. The publishers are early because they had to be, and their architecture, layered detectors plus mandatory human adjudication, is the template the rest of the document economy will be forced to adopt.
What Enterprise Buyers Should Take Away
For CIOs and CTOs, the MDPI deployment is less interesting as a publishing story than as a reference architecture for content authenticity at volume. The components are instructive: a continuous monitoring layer that watches an item across its entire lifecycle, multiple specialized detectors covering text, images, and behavioral signals, risk scoring that triages rather than decides, and a human review tier with the final say. That shape generalizes cleanly to fraud detection, KYC document review, and any pipeline where adversaries are actively producing synthetic inputs. Buyers evaluating integrity tools should ask vendors how their products compose with others, because no single model will hold the line.
The harder organizational question is governance. MDPI's insistence that AI generates signals while humans make decisions is not a technical detail; it is the policy that determines liability, fairness, and trust. Enterprises rushing to automate verification should write that boundary into procurement and process now, before a false positive damages a customer, a candidate, or an employee. The race Sayas describes is real, and it will not end. The winners will be the organizations that treat AI integrity screening as a standing operational capability with human accountability built in, rather than a one-time tool purchase that quietly degrades as the adversaries adapt.
