Welcome to the First SIA Special Report

This series examines how AI is reshaping scholarly publishing and what that means for research integrity and the scholarly record. In this installment, we focus on the challenges and risks of AI overreliance across publishing workflows, from editorial processes to attribution, drawing on original interviews, public discussions, and published commentary from individuals working within institutions under pressure, both on platforms and in editorial roles.

What This Report Covers

For this first part, we conducted original interviews with Ramin Zabih, Executive Director of the open-access archive arXiv, and Ben Williamson, Editor of Learning, Media and Technology (Taylor & Francis). Additional perspectives are drawn from practitioners and researchers who have written and spoken publicly on these issues. We hope this provides a useful overview of current attitudes towards AI adoption, the questions still to be addressed, and some lessons learned thus far. As the landscape continues to shift rapidly, we believe these perspectives offer a timely and grounded starting point for the broader conversation.

How Current Publishing Workflows Are Adapting to AI

The European University Association recently identified the rise of AI as one of the key factors shaping the evolution of the scholarly communication system. In a webinar they organized to explore these issues at the end of last year, most attendees still felt too uncertain to determine how AI would impact open science policies at their institutions.

Although some institutions are already using AI tools in their daily operations, not all have invested time in training their researchers to do so. Most publishers focus solely on demanding declarations on the use of AI tools, even as their use in research and manuscript writing keeps evolving. See the case of IOPScience.

Platforms, Preprints, and the Limits of Trust

A key example of what is at stake is Open Evidence. This organization has licensed journal content from publishers, built a clinical AI interface around it, and attracted a critical mass of physicians. This is not just building AI tools; they are aggregating both the content and the audience into a single platform, raising capital at a $12 billion valuation. What publishers can license in an AI economy, however, is not just information.

Angela Cochran & Todd Toler, The Scholarly Kitchen, March 2026

And it is not just traditional publishers that need to address these questions. In the past, submissions of review, survey, or position papers to arXiv were extremely rare. This all changed in 2025, when moderators realized that over 10% of all submissions to the Computer Sciences (CS) category were of very low quality, illustrating the scale of the problem that platforms built on trust, volunteer work, and a sense of community are facing. We spoke with Ramin Zabih, Executive Director at arXiv, who told us that the priority is to ensure a fair allocation of resources; thus, arXiv seems willing to forego being at the forefront of timely access to survey data, which quickly becomes outdated. The surge in low-quality submissions has had a direct human cost, with moderators seeing an estimated two- to fourfold increase in workload. As Zabih notes, this level of pressure is not sustainable. When faced with large-scale automated submissions, the need to rapidly address these issues prompted a policy change in October 2025, restricting review and position papers in the CS category. Is this the end of the dream to accelerate science that was the inception for the original preprint platform? Others have yet to determine their policies, and this is often hampered by structural limitations.

Ramin Zabih, Executive Director at arXiv

For platforms like SocArXiv, the open archive of the Social Sciences, demographer Philip N Cohen concedes that the resources to verify claims, sleuth out people who make false claims, or deny using chatbots when they actually do are absent. In a follow-up exchange, Zabih told us that arXiv is actively developing a range of tools to detect AI-generated content, spanning automated screening, metadata checks, moderation support, and more. He was candid, however, that effectiveness varies as submitters adapt their tactics in response.

The case for AI in peer review

The October–December 2025 issue featured perspectives from Daniel Acuña (Reviewer Zero AI), Mario Malicki (Stanford SPORR), Paul Whaley (Evidence-Based Toxicology), and Daniel Ucko (American Physical Society) on how AI might help address the growing strain on peer review.

Read the Article

The Zombie Citation Problem

Publishers have always held authors accountable for the correctness of content in their submissions, and this includes the reference list. The term "zombie citation" was once used to characterize the persistence of citations to retracted papers in novel scientific publications, but has expanded to include "hallucinated" citations. When Ben Williamson, Editor of Learning, Media and Technology, encountered a new submission citing a paper with him as author that he had never written, he embarked on a chase to track down the origin of that citation over time. But the sheer volume of submissions with problematic reference lists has increased over time, and with it the realization that checking each reference at the editorial level seems impossible.

At the Prophy Predicts webinar on future research integrity threats held in February 2026, Jason Hu, Director of Research Integrity Engagement at Taylor & Francis Group, noted that in the last few years every publisher has seen double-digit growth in submissions, but not necessarily in acceptance rates. Does this mean that peer reviewers are being exposed to ever increasing quantities of poor-quality research? Let’s not forget that peer reviewers' work is still mostly unpaid and invisible.

In a recent blog post, Marco Marabelli, associate professor at Bentley University, poses some key questions: in future, will journals and publishers with “stricter” GenAI policies be viewed as being more legitimate or credible, and will GenAI use affect paper acceptance and impact?

Ben Williamson, Editor, Learning, Media and Technology

Misuse of Generative AI Tools

It is crucial to distinguish other automated workflows and processes from generative AI, data systems that use machine learning algorithms to measure and make predictions about behaviors or outcomes, and which include large language models that produce original text. Misuse of generative AI tools seems to be amplifying broader structural incentives in academia, such as pressure to publish quickly or frequently. But the intrinsic limitations of these novel tools are often not fully apprehended even by experienced academic users. Those who seek training do not always find it.

One of the clearest and in-depth examples we saw is the novel institutional guidance at Oxford University, which states it is important to “consider the probabilistic nature of GenAI tools, which includes the tendency to produce a different output from the same input”. This policy focuses on judicious choice of any AI tool considering its fitness for purpose and does not make blanket statements covering all tools, which would preclude them from use to generate knowledge.

Accountability, Governance, and Geopolitical Pressures

In January 2026, arXiv altered its endorsement requirement for first-time submitters to the platform. Perhaps requiring evidence of prior authorship or a personal endorsement from an established arXiv author could help restore accountability to the community being served. Nonetheless, in a follow-up exchange, Zabih told us that arXiv is developing tools to detect AI-generated content, including automated screening systems, metadata checks, moderation support tools, and more. The shared-governance challenge is obvious, and placing blame on a single stakeholder group seems unreasonable at best. Ben Williamson argues that global technology companies have begun acting as governance organizations in education, potentially amplifying existing bias and discrimination, and causing the accountability issue to become enmeshed in the current political climate. For Williamson, AI companies are not merely offering tools; they are becoming the infrastructure through which academic work and scholarly inquiry take place, reshaping the conditions of knowledge production in ways that remain largely ungoverned.

Also speaking at the Prophy Predicts webinar, Jayne Marks, CEO of Maverick Publishing Consultants, noted that this is further amplified by the fact that different geopolitical regions have different policies towards access and open science, which can lead to skewing of publishing in favor of or away from different regions. She also cautioned about the rapidly increasing ability of bad actors, such as paper mills, to hack into core shared infrastructure such as CrossRef, warning that this could call into question the reliability of the entire system.

Graham Kendall, Vice Chancellor, MILA University Malaysia

The question of who has the authority to act on these threats remains largely unresolved. Fellow panelist Graham Kendall, Deputy Vice Chancellor at MILA University Malaysia, argued that scholarly publishing, a multi-billion dollar industry, needs an international body with real power to act, one that researchers and authors can consult to verify whether a publisher meets minimum standards. Without it, he warned, bad actors face no meaningful consequences.

The new issues that have arisen regarding transparency in decision-making are hard to address with our current knowledge. Is training in this novel technology desperately lacking, or are the institutions responsible for training us adopting unverified processes? AI has been used by educational institutions and governance bodies to analyze student data, and companies have integrated AI features into products that schools were already using. While these tools may contain factual inaccuracies, they also rest on a narrow assumption about what constitutes learning. At the individual level, overreliance on these tools has already been linked to deskilling and to the impairment of critical thinking. Will the use of AI tools start to narrow the focus of the research being conducted?

Paul Whaley, Editor-in-Chief, Evidence-Based Toxicology

Use of LLMs by LX English Writers

While AI assistance in academic writing has surged indiscriminately across journals, a recent study found that researchers from non-English-speaking countries are more likely to rely on AI writing tools than native English speakers. This also seems to be observed in academic institutions, with Oxford University stating that its policy does not apply in cases of GenAI tools being used for tasks such as assisting non-native speakers with translation, transforming transcribed spoken language to written language, formatting documents, or improving one's own standard of English or a foreign language.

In parallel, arXiv’s updated non-English paper policy, which took effect in February 2026, was meant to increase the number of non-English submissions. The announcement was met with tremendous backlash and even accusations of discrimination. Some machine translations are still deemed to be of very poor quality, especially in academic writing, raising questions about whether arXiv and other venues are considering mitigation strategies to avoid unintentionally narrowing access to academic publishing. Zabih is straightforward about the dilemma: preserving the platform means making it harder to submit low-quality AI-generated content, which will inevitably, and accidentally, also raise barriers for some legitimate users. "We are doing our best to mitigate this side effect," he told us, adding that arXiv is making alternative submission paths available. In the current situation, he notes, there are no great options.

Research and Collective Action

Research on these issues is ongoing. An initiative called Taming the Crocodile united an unprecedented group of publishers in a study led by Kudos to examine the impact of zero-click search and AI-generated overviews in scholarly communications. In addition, Cochrane has launched a study to test whether AI tools can support or enhance evidence synthesis (see Milestones, pp. 34–39).

As linguistic bias is still intrinsically linked to geographic and ethnic disparities, how can this chance for deep structural reform be used to resolve asymmetries rather than amplify them? Collaboration was recently hailed as the force behind capacity building among scholarly publishing stakeholders, and the basis for the creation of United2Act. It has also shaped thinking that aims to flag integrity anomalies much earlier, thereby reducing the burden on publishers with respect to retractions.

Jayne Marks, CEO, Maverick Publishing Consultants

As research and higher education become ever more dependent on AI infrastructure, is data being confused with knowledge? Will the near future be fairer, or will the erosion of trust in science be widespread?

Subscribe to read our full interviews with Ramin Zabih and Ben Williamson in SIA's expert-led forum. Have a view from inside your own platform, journal or institution? Join the conversation.