Image: revistapesquisa.fapesp.br · rights & removal

The miracle of multiplying bibliographic references (ghosts)

Reporting by Pesquisa FAPESPRead the original at revistapesquisa.fapesp.br

Executive Summary

The proportion of biomedical articles containing at least one fabricated bibliographic reference increased significantly between 2023 and early 2026, growing from approximately 4 per 10 thousand in 2023 to 56.9 per 10 thousand by the beginning of 2026. This finding stems from an audit of 2.5 million published studies from 2023 to February 2026, which aggregated over 125 million references. The acceleration in this trend is associated with the widespread adoption of Large Language Models (LLMs) and changes in journal indexing practices. An automated system was developed to verify bibliographic records against web-available articles. Recommendations proposed include integrating automated reference verification before peer review and conducting retrospective screenings of journals for necessary corrections, suggesting retractions for manuscripts with false references. Concerns exist regarding the role of AI in peer review, including potential reinforcement of unequal knowledge distribution, and a debate over where the line should be drawn between human and artificial oversight.

Facts Only

* In 2023, approximately 4 in every 10 thousand biomedical articles contained at least one fabricated bibliographic reference.
* The proportion increased to 51.3 per 10 thousand by the last quarter of 2025.
* The proportion reached 56.9 per 10 thousand at the beginning of 2026.
* The audit gathered 2.5 million published studies from 2023 to February 2026.
* The manuscripts totaled over 125,615,773 references.
* An automated system compared bibliographic records with true articles on the web for verification.
* The acceleration coincided with the mass use of LLMs like ChatGPT and Claude between the end of 2023 and the beginning of 2024.
* Researchers estimate the time between article submission/publication and the rise in manipulated references was ten to two hundred days later.
* Some surveys suggest more than half of reviewers resort to artificial intelligence at some stage of the review process.
* A pilot project by NEJM AI initiated a rapid approval model based on AI analysis followed by human editor review.

Full Take

The narrative highlights a systemic challenge where technological advancement, specifically LLMs, interacts with established scientific publication processes to create new integrity risks. The acceleration in fabricated references is contextualized not merely as an artifact of poor ethics but as an outcome of shifts in scientific incentives—namely, the focus on article production over discovery fragmentation—which predate the AI influence. The development of automated verification and proposals for retraction suggest a move toward external, mechanical enforcement of veracity, driven by accountability mechanisms like exemplary punishment. The tension lies between using AI to detect errors versus the fear that AI itself could be used to create more sophisticated forms of exclusion in knowledge dissemination, potentially amplifying existing geopolitical inequalities within science. The assertion that the peer review crisis precedes AI development suggests that technological tools amplify pre-existing structural pressures rather than being the primary cause. The central unresolved question is determining the appropriate boundary and integration level for artificial intelligence in maintaining scientific rigor.

From the original · Pesquisa FAPESP

In 2023, about 4 in every 10 thousand biomedical articles contained at least one fabricated bibliographic reference. It doesn't seem so at first glance, but look at the growth rate: in the last quarter of 2025, the proportion had already jumped to 51.3 per 10 thousand and, at the beginning of 2026, it reached 56.9—about 14 times more.

Read the full story at revistapesquisa.fapesp.br

Sentinel — provisional

Raw score, 0 to 1 (uncalibrated)0.10

No strong signs of machine writing were found in the source article. Provisional estimate, not a finding that a person wrote it.

The model's own note (unverified)

This text is a high-quality synthesis of recent research on AI's impact on scientific publishing, skillfully linking empirical data with philosophical discussions about peer review and knowledge production.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

What the model noticed in the source article

Erratic sentence rhythm and nuanced flow mixed with precise data presentation.

Passionate exploration of complex, interlocking themes (AI, publishing ethics, review processes) without the detached balance typical of pure synthesis.

Logical progression moving from specific data (references) to broader philosophical implications (peer review crisis) with intentional counter-arguments.

Citations of specific audits, names (Bauchner, Rivara), and journal reports suggest grounding in real-world research reporting.

Signs the model read as against machine writing

The piece features an idiosyncratic shift in tone ('Fire against fire?', 'spoken with affection, dear reader') which suggests a human voice engaging the material rather than simply relaying it.

The complexity in weaving together specific academic findings (like the 14x growth rate) with broader ethical commentary points toward specialized knowledge synthesis.

From the original · Pesquisa FAPESP

In 2023, about 4 in every 10 thousand biomedical articles contained at least one fabricated bibliographic reference. It doesn't seem so at first glance, but look at the growth rate: in the last quarter of 2025, the proportion had already jumped to 51.3 per 10 thousand and, at the beginning of 2026, it reached 56.9—about 14 times more.

Read the full story at revistapesquisa.fapesp.br