Among the challenges emerging with artificial intelligence, verification is undermined at both ends of the scholarly publishing pipeline. Authors are now able to easily create academic work with plausible but fake references. Reviewers, also increasingly helped by the same technology, use tools that cannot detect them. The result is a closed loop where fluency stands in for truth. This dynamic is described in two recent publications. The Lancet’s audit of 2.5 million papers found around 4,000 fabricated references in more than 2,000 articles, a rate that has increased more than twelvefold between 2023 and 2026 [1]. Most alarming of all: 98.4% of affected papers saw no publisher action [1]. A parallel scoping review describes the reviewer side of the equation. AI-generated reviews that are “superficial, overly positive, and general” fail to detect fabricated data or methodological inconsistency, and are significantly more lenient than human reviewers [2]. AI-human agreement sits well below human-human reliability, and consequently, an AI-generated review will let a fabricated citation pass [2].
Beyond being an academic housekeeping problem, this poses a clear risk to patient safety. A faulty primary reference propagates into guidelines, formulary decisions, and treatment pathways. The scoping review states the point explicitly: flawed medical review “can propagate directly into clinical practice guidelines... and patient harm at scale” [2]. As the evidence base is quietly contaminated, the damage may surface years later at the bedside.
Examining the current practices among top publishers can help streamline the actual existing gaps. A 2025 cross-sectional analysis found that 78% of the top 100 medical journals had AI peer-review guidance by mid-2024, and 59% explicitly prohibited reviewer use [2]. Despite these regulations, reviewer AI use remains widespread in practice, and policy hardly ensures enforcement [2]. On the author side, AI is broadly permitted with disclosure and author accountability under COPE and ICMJE frameworks. AI use is more restricted on the reviewer side with confidentiality being the central concern [3][4]. Policies still vary largely by publisher and by journal. For instance, Elsevier imposed strict prohibitions, while Springer Nature and Wiley permit limited use with disclosure and confidentiality safeguards [5][6][7]. On reference checking specifically, almost no publisher runs systematic automated verification pre-review, which is precisely the gap the Lancet audit exposes [1]. Detection remains weak, where generic AI-text detectors suffer high false-positive rates, especially against non-native English authors. Meanwhile, domain-specific methods such as token signatures and context anchoring are promising but not yet standard [2].
The promising news is that frontier tools exist, and they fall into three broad categories: mature verification pipelines, emerging editorial aids, and governance mechanisms. The most mature tool is reference verification such as the Lancet pipeline, which achieved precision around 91% [1]. These pipelines check citations across multiple sources such as PubMed, Crossref, OpenAlex, and Google Scholar, then combine text-similarity scoring with LLM screening to distinguish format errors from outright fabrication [1]. Alongside this, agentic reference-integrity frameworks are beginning to emerge in the information-science literature [8]. A second category covers editorial triage where AI can handle plagiarism, reporting-guideline, and format checks. Considerably, reviewer-matching recommenders already outperform manual search [2]. A third category concerns oversight and accountability. Auditable prompt logs and watermarks for human-AI interactions in editorial workflows have been proposed as a medium-term reform [2]. Separately, bias-monitoring audits of AI scoring tools, checking for affiliation, gender, and linguistic bias, should become routine before deployment. Anomaly detection can also serve as a first-line filter, provided that every flag receives human review. Overall, the evidence indicates that AI strengthens review as a transparent support tool under human oversight, and only as an assistive mechanism [2].
Several concrete policy directions are worth pursuing, ranging from changes individual publishers can make to reforms that require collective action. Role-specific AI policies that differentiate author, reviewer, and editor contexts stand out as the scoping review’s key recommendation [2]. Integrity metadata in indexing records is a second recommendation, where downstream users can see whether a paper’s references were verified. Additionally, research-integrity databases such as COPE still lack a discrete “fabricated reference” category, and adding one would make the problem visible [2]. Cross-publisher integrity databases and shared governance standards represent the scoping review’s long-term recommendation, and these are now emerging through COPE/ICMJE/WAME coordination [3][4][9]. Regulation is pushing in the same direction. HIPAA and GDPR constraints make uploading manuscripts to third-party LLMs a compliance risk, which is already nudging publishers toward secure, auditable infrastructure [2].
The tools exist, and the policies are within reach. What remains is the decision to use them before the next contaminated reference list becomes a clinical guideline. The question is no longer whether publishing has a verification problem, it is whether it has the will to verify.