Citation manipulation may not be the most visible research integrity issue, but it is rapidly becoming one of the most consequential. In seconds, large-language models can generate references that look convincing but often turn out to be problematic, referring to either completely fabricated references (increasingly known as hallucinations) or valid sources that don’t actually support the claims they’re linked to. At the same time, citations continue to serve as a key form of recognition, influencing careers, funding decisions and reputations. As a result, the incentive to boost or manipulate citation counts continues. Seen in this wider context, citation manipulation is creating a fast-growing challenge for research integrity. 

Although the exact scale of the problem is disputed, a recent paper in the Lancet that looked at 2.5 million biomedical and life sciences papers published between 2023 and 2026, states that in 2023, approximately one in 2,967 papers contained at least one fabricated reference. By 2025, this had risen to one in 458 and in the first seven weeks of 2026, one in 277 papers had at least one fabricated reference. If the references are fabricated, no one benefits beyond the author, who we have to assume is taking a short-cut to publication by using an LLM to help situate their work in context (let’s be generous with our assumptions). Many people’s time is wasted, however: editors, reviewers, and, most important, readers, who trust what we publish and build their work on our content. 

As generative AI becomes more accessible and widely used, it may also be driving papermills to diversify, as authors no longer need to rely on paid writing services. Increasingly we see mills pivoting to offer citation services, with neatly tiered packages of 10, 25 or 50 guaranteed citations within a given timeframe to boost their H-index. These can be placed in articles that the mills still produce or inserted by unscrupulous editors and reviewers working on behalf of the mill. If the purpose of the publication is citation hosting, it may explain why we sometimes see articles (and preprints) with totally fabricated author lists. Now we’re seeing a whole range of different people benefitting, and the same people’s time being wasted. 

Citation mills are inherently more difficult to identify than hallucinated references. They often appear, at first glance, consistent with normal scholarly activity. As a result, the challenge for publishers and editors is not only to detect anomalies, but to interpret intent whilst working with little evidence. Some citation styles also make spotting any trends difficult (et al, anyone?!). 

Citation manipulation often sits in a difficult grey area, where intent and impact are not always easy to untangle. For instance, an author’s name could be misspelled, a link could be broken, metadata may be incomplete. Any other simple mistake might have crept in. But what’s a genuine error, and what’s an attempt to manipulate? How do you decide what to fix and what to investigate? 

For example, excessive self-citation or the inclusion of irrelevant references may suggest an attempt to inflate metrics. However, even these signals are not always straightforward. In smaller or highly specialised fields, high levels of self-citation can be a natural consequence of a small research community. Similarly, references that initially appear irrelevant may in fact point to related methodologies, similar applications or simply reflect unclear or misleading titles. 

Complicating matters further are the power dynamics within the publication process. Authors may feel obliged to incorporate citation suggestions from reviewers or editors, even when they consider them unnecessary. As a result, citation patterns may reflect compliance rather than deliberate manipulation. This makes citation manipulation particularly challenging to identify and address, as the available signals are often ambiguous and highly dependent on context. 

It might seem like something that technology can fix. Detection tools are indeed improving, and many publishers are now introducing checks during submission to flag suspicious references earlier. But determining what constitutes fraud or error turns out to be a nuanced and time-consuming issue (Marie Souliere or Ioana Cristea), so human judgement still plays a crucial role. At IOP Publishing we rarely see direct evidence of manipulation or intent, such as an agreement or transaction. Instead, we’re left with patterns that look concerning but are hard to prove definitively. 

This creates a dilemma for editors, who must assess context and fall somewhere between supporting honest authors who have made honest mistakes and penalising intentional manipulation of the scholarly record. As a result, citation manipulation has historically been treated inconsistently. 

Recognising this evolving issue, the Committee on Publication Ethics (COPE) has updated its  to explicitly include citation manipulation as a form of compromised publication practice. In the significant 2025 update, the guidance states retraction is appropriate where “the peer review or publication process was compromised” (eg., fake reviewers, paper mill use, or citation manipulation). The guidance is designed to provide editors with clearer language and greater confidence in addressing such cases. to explicitly include citation manipulation as a form of compromised publication practice. In the significant 2025 update, the guidance states retraction is appropriate where “the peer review or publication process was compromised” (eg., fake reviewers, paper mill use, or citation manipulation). The guidance is designed to provide editors with clearer language and greater confidence in addressing such cases. 

Other stakeholders have also stepped in. Recently the physical sciences repository arXiv announced that it will ban researchers from posting their manuscripts on the platform for one year if a submission is found to contain references that have been hallucinated by artificial intelligence tools. The ban also applies to papers showing other clear signs of unverified generative AI use, where outputs have not been properly checked. Such “incontrovertible” evidence can include fabricated references or comments generated during interactions with large language models. 

Publishers (including IOP Publishing) are also adapting their policies and workflows. We have introduced submission-stage checks for hallucinated references, using Alchemist Review as part of a broader effort to identify issues earlier in the process. However, these checks still rely on human verification and citation investigations cannot be reduced to pattern matching alone. Editors and research integrity teams must apply contextual judgement by looking at whether a citation exists, whether it supports the claim being made, and whether emerging patterns point to carelessness or deliberate manipulation. The challenge is not simply identifying flawed metadata, but interpreting both intent and impact. And of course, every questioned author pleads a “carelessness” defence. 

Because the underlying pressures in the research system remain unchanged (citations still serve as a key measure of impact, shaping reputations and influencing funding decisions), the issue is unlikely to go away.  

However, there is growing recognition that this problem calls for shared standards across the publishing community including institutions, funders and publishers. Citation manipulation sits at the intersection of incentives (how research is evaluated), technology (how research is produced) and judgement (how research is selected for publication). The evidence, however, is often complex and messy and the tools available are not infallible. This is precisely why technology (supported by human assessment), clear policies and stronger shared guidance are becoming more important. However varied our approaches, we all seem to agree on one thing: ultimate responsibility rests with authors, who must read, verify and take full accountability for every source they cite.  

Citations may seem a small part of academic output, but they underpin credible, reliable research. They enable others to build on existing ideas and support discoveries that shape entire fields and the way we understand the world and ourselves. But like humans, they are messy, varied and there are multiple systems and drivers involved. Accepting this is a knotty problem that will require multiple solutions, we as a community need to be clear about what we expect and what we stand for. We need authors to be supported in pushing back on citation suggestions. We need clear policies. Great tech to identify problems at scale and reviewer and editor education. Safe whistleblower mechanisms. We need consequences for those who are found to have manipulated the process. We need to share data on offenders. All of this will help with detection and prevention to an extent. But to really make it go away? We need structural incentive reform.