This is a recap of the Prophy Predicts webinar — "Future Research Integrity Threats" — hosted on February 19, 2026. The full recording is available on YouTube.
Research integrity threats are evolving. And the scholarly publishing industry is not keeping pace.
That was the unspoken consensus running through our first Prophy Predicts panel — a 60-minute conversation with four people who spend their working lives on opposite ends of the problem: a publishing consultant, a major publisher, a vice chancellor, and a production workflow expert.
We asked each panelist the same question ahead of the session:
What new or unexpected research integrity threat are we likely to see in the next two to five years?
What came back was sharper than we anticipated.
Prophy Predicts: Future Research Integrity Threats brought together four panelists for a moderated discussion on what's coming next for scholarly publishers, researchers, and the broader research ecosystem.
Panelists:
Moderated by Gareth Dyke, Partnership Director at Prophy.
Jayne Marks opened with three predictions. The first two — geopolitical fragmentation of publishing access and AI narrowing researchers' field of view — are concerns the industry is already wrestling with.
Her third one is not widely discussed.
"This is the one that really keeps me up at night," she said. "What happens if paper mill companies hack into some of our core infrastructure?"
She named Crossref specifically — the DOI registration and metadata backbone that much of scholarly publishing depends on. If bad actors gained access and started manipulating citations, author names, or DOIs, the downstream effect would be hard to contain and even harder to detect quickly.
It is not a hypothetical. It is a logical extension of how organized bad actors already operate. The publishing industry has grown used to thinking about fraud at the article level. Jayne's warning is that the attack surface is bigger.
Graham Kendall, Vice Chancellor of MILA University and a researcher with over two decades of AI experience, named the combination of AI and paper mills as the most serious near-term threat to research integrity.
His reasoning goes in steps.
First: AI makes it trivially easy to generate manuscripts at scale. He pointed out that in a room of 30 people on the webinar call, the group could generate hundreds of plausible-looking papers and probably get half of them published. Paper mills already operate this way — generating content, selling authorships, and exploiting corrupt editors to get papers into the record.
Second: those papers, once published, enter the scientific literature. They contain hallucinated references, unsupported hypotheses, and fabricated results. Retractions happen, but slowly, and many problematic papers are never caught.
Third — and this is the part that unsettled the panel — those bad papers then become training data for LLMs.
"In two to five years' time, when you ask an LLM a question, it's going to draw on stuff that it wrote two years ago, which is nonsense anyway," Kendall said. "AI is writing papers which are nonsense, and then LLMs are learning from those papers."
The loop closes. Bad papers train AI systems. Those AI systems generate new bad content. The scientific record degrades from within.
He also flagged citation mills as a parallel and underdetected problem — circular citation networks designed to inflate H-indices. He described a recent case of a researcher who published 124 papers in a single year, mostly self-citations, and built an H-index of 24 doing so. The metadata infrastructure needed to reliably detect these patterns simply does not exist in a usable form yet.
Jason Hu, who has spent ten years in scientific publishing at Taylor & Francis, framed the threat differently.
"My biggest worry is that we are shifting from fake papers to fake knowledge systems," he said.
What he means: publishers used to worry about individual fraudulent articles. The problem now operates at a different level. Fake citations get inserted to boost institutional rankings. Retracted papers continue to be cited because the rest of the literature hasn't been updated. Readers stop engaging with primary sources and turn to ChatGPT for summaries instead.
Each of these individually is manageable. Together, they describe a scientific knowledge infrastructure that is quietly losing reliability.
Jason also pointed to undisclosed AI use in peer review as a growing problem. The consensus among publishers and funders is that reviewers should not use AI to write their review comments without disclosure. The reality is that many do. "It's eroding the trust between reviewers, editors, and authors, but also eroding the creativity of human beings," he said.
The submission volume numbers back up his concern. Publishers across the board are seeing double-digit growth in submissions. Acceptance rates are not growing at the same pace. That gap represents a volume of AI-assisted, low-quality, or entirely fabricated content pressing against peer review systems that were never designed to filter at this scale.
He called for a shift in how researchers are evaluated — away from "how many papers in which journals" and toward the quality and process of the research itself. He acknowledged it will be very hard to implement. "The previous evaluation metrics are not going to work, at all."
Sukhwant Singh, COO of Newgen KnowledgeWorks, brought the workflow perspective.
His term for what's coming: credibility dilution.
"There may be many papers that are not fake science, not misconduct. But because of the sheer volume, there's a credibility dilution — it becomes harder for editors, reviewers, and readers to assess the reliability and robustness of the research."
This is a different kind of problem from the ones the industry's current tools are built to catch. Plagiarism checkers and image manipulation detectors look for specific, identifiable signals. Credibility dilution doesn't announce itself. It's a quiet degradation in the overall signal-to-noise ratio of the literature.
Research content, Sukhwant argued, is following something like Moore's Law — volume growing exponentially. The workflows built to handle it are not keeping up. "Traditional workflows were not designed to handle this kind of volume, and we actually see the impact on the ground."
His prescription: move the focus upstream. Publishing workflows were built around detection after a manuscript enters the system — checking for plagiarism, flagging authorship concerns, screening for image manipulation. These checks happen too late, and at too high a volume to be reliable.
"The shift needs to move from detecting bad actors at the point of publication to a point where it is being submitted, or even before that. We need mechanisms that strengthen the research earlier in the workflow."
On paper mill detection specifically, he was direct: there is no single reliable mechanism right now. The signals are distributed — metadata inconsistencies, submission behavior patterns, structural anomalies across manuscripts, mismatches between claims and data. Catching them requires combining AI-based screening with human judgment, and integrating that combination into the workflow much earlier than it currently sits.
As the session moved to closing thoughts, the conversation circled back to a structural problem that underpins nearly all the threats raised.
Jayne put it plainly: "We've got all these perverse incentives. The incentives from funders to institutions, from institutions to researchers, from researchers to publishers — none of it works together."
Research managers in institutions, she noted from her own research, are told not to use H-factors to evaluate researchers. So they stop using them officially. Then they use them anyway, because there's nothing else. "I'm sure funders would be horrified to think that we were all still chasing impact factors."
Graham added the enforcement dimension. In law, medicine, and architecture, misconduct leads to being struck off. In academic publishing, consequences are limited. He cited the Omics case: a publisher fined $50.1 million for unethical publishing practices. They never paid. They're still operating.
"We need an international body with teeth that can actually take control of this multi-billion dollar industry," he said. "People are still using Beall's List from 2012, which is ten years out of date."
Jason framed it as a call for a different kind of collaboration. The United2Act initiative — sponsored by COPE and the STM Association — is one example of stakeholders coming together across publisher, funder, researcher, and service provider lines to address paper mill risks specifically. He chairs the education working group.
"We serve the same purpose. We're here to foster human growth through new knowledge. We need to try very hard to be neutral, find the problem together, and talk about how to resolve it."
If you're a Peer Review Manager, Head of Peer Review, or Editor-in-Chief, the practical takeaway from this session is not a list of new tools to acquire. It's a shift in how you think about where integrity risk enters the workflow.
The current model — submit, review, screen, publish — was designed for a different volume and a different threat environment. The threats described in this session are not fringe risks. They are already present at scale, and they are getting harder to detect with tools and processes built for a slower world.
The panelists converged on a few specific directions worth acting on now.
Moving pre-submission checks earlier. Detecting structural and behavioral anomalies at the point of submission, not after a paper has moved through peer review. This is where Sukhwant's workflow argument is most actionable: the order of the process matters, not just the tools within it.
Building citation pattern analysis into editorial workflows. Citation mills are already operating. The metadata infrastructure to track them reliably doesn't exist yet, but the behavioral patterns are visible to anyone looking systematically across submissions.
Treating cross-stakeholder collaboration as infrastructure, not a nice-to-have. No single publisher, tool, or policy solves this alone. Researchers, publishers, funders, and service providers running separate responses to the same threat are giving bad actors the space to operate in the gaps.
Prophy Predicts is an ongoing webinar series bringing together researchers, publishers, and industry experts to discuss the future of scholarly publishing and research integrity. The next event is scheduled for March 2026.
Prophy helps publishers protect research integrity at scale — through reviewer discovery, conflict of interest detection, and early warning systems that analyze patterns across 180+ million articles.
If you work in peer review operations or research integrity and want to discuss how your current workflow holds up against the threats raised in this session, get in touch with our team at prophy.ai.
This recap was produced by the Prophy team from the full session recording. Quotes have been lightly edited for clarity.