Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals
The critical vulnerability of the traditional blind peer-review system to the challenges posed by the rapid development of generative tools is examined. Based on a real-world precedent involving the discovery of a completely fabricated bibliography in a submitted manuscript, we analyse the basic mec...
Saved in:
| Date: | 2026 |
|---|---|
| Main Authors: | , , |
| Format: | Article |
| Language: | English |
| Published: |
Publishing House "Akademperiodyka" of the National Academy of Sciences of Ukraine
2026
|
| Online Access: | https://ojs.akademperiodyka.org.ua/index.php/Zoodiversity/article/view/971 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | Zoodiversity |
| Download file: |  |
Institution
Zoodiversity| _version_ | 1869381616028090368 |
|---|---|
| author | Kharchenko, V. O. Korneyev, V. O. Filimonova, N. S. |
| author_facet | Kharchenko, V. O. Korneyev, V. O. Filimonova, N. S. |
| author_institution_txt_mv | [
{
"author": "V. O. Kharchenko",
"institution": "I. I. Schmalhausen Institute of Zoology NAS of Ukraine"
},
{
"author": "V. O. Korneyev",
"institution": "I. I.Schmalhausen Institute of Zoology of the NAS of Ukraine"
},
{
"author": "N. S. Filimonova",
"institution": "I. I. Schmalhausen Institute of Zoology NAS of Ukraine"
}
] |
| author_sort | Kharchenko, V. O. |
| baseUrl_str | https://ojs.akademperiodyka.org.ua/index.php/Zoodiversity/oai |
| collection | OJS |
| datestamp_date | 2026-06-29T16:20:36Z |
| description | The critical vulnerability of the traditional blind peer-review system to the challenges posed by the rapid development of generative tools is examined. Based on a real-world precedent involving the discovery of a completely fabricated bibliography in a submitted manuscript, we analyse the basic mechanisms behind the creation of false references, such as anachronisms, “Frankensteinisation”, and professional biases. The study demonstrates the evolution of the threat: a transition from the obvious errors of early algorithms to the deep “semantic hallucinations” of modern RAG-based search engines, which are capable of generating perfectly formatted yet conceptually empty texts derived from real databases. To protect the publication process, an updated algorithm for editorial control is proposed, requiring the mandatory validation of digital object identifiers (DOIs) and a clear declaration of the algorithms utilised by the authors. The main conclusion emphasises the necessary and unalterable transition to the Open Science Framework paradigm, where textual material is viewed merely as an accompanying document to a verified array of primary datasets, open-source code, and deposited collection specimens. |
| doi_str_mv | 10.15407/zoo2026.03.301 |
| first_indexed | 2026-06-30T01:00:34Z |
| format | Article |
| fulltext |
DOI 10.15407/zoo2026.03.301
UDC 001.81:004.8:050
PHANTOM REFERENCES AND THE PEER-REVIEW CRISIS:
HOW ARTIFICIAL INTELLIGENCE TESTS
THE RESILIENCE OF SCIENTIFIC PERIODICALS
V. O. Kharchenko, V. O. Korneyev *, N. S. Filimonova
Zoodiversity Journal Editorial Board
* Corresponding author
E-mail: valery.korneyev@gmail.com
V. O. Kharchenko (https://orcid.org/0000-0002-3824-2078)
V. O. Korneyev (https://orcid.org/0000-0001-9631-1038)
Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of
scientific periodicals. Kharchenko, V. O., Korneyev, V. O., Filimonova, N. S. — The critical vulner-
ability of the traditional blind peer-review system to the challenges posed by the rapid development of
generative tools is examined. Based on a real-world precedent involving the discovery of a completely
fabricated bibliography in a submitted manuscript, we analyse the basic mechanisms behind the cre-
ation of false references, such as anachronisms, “Frankensteinisation”, and professional biases. The
study demonstrates the evolution of the threat: a transition from the obvious errors of early algorithms
to the deep “semantic hallucinations” of modern RAG-based search engines, which are capable of gen-
erating perfectly formatted yet conceptually empty texts derived from real databases. To protect the
publication process, an updated algorithm for editorial control is proposed, requiring the mandatory
validation of digital object identifiers (DOIs) and a clear declaration of the algorithms utilised by the
authors. The main conclusion emphasises the necessary and unalterable transition to the Open Sci-
ence Framework paradigm, where textual material is viewed merely as an accompanying document to
a verified array of primary datasets, open-source code, and deposited collection specimens.
Key words : academic integrity; LLM; generative hallucinations; open science; data deposition;
research falsification.
Introduction
The rapid development of large language models (LLMs) has created an unprece-
dented challenge for academic integrity. While artificial intelligence initially dis-
pelled fears by acting as a convenient tool for improving academic English, editorial
Editorial Zoodiversity, 60(3): 301–308, 2026
© Publisher Publishing House “Akademperiodyka” of the NAS of Ukraine, 2026. The article is
published under an open access license CC BY-NC-ND (https://creativecommons.org/licenses/
by-nc-nd/4.0/)
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
V. O. Kharchenko, V. O. Korneyev & N. S. Filimonova
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
302
boards of scientific journals today face a new, far more dangerous trend — the com-
plete fabrication of scientific data and bibliography. The propensity of LLMs to gen-
erate plausible-sounding but fictitious references was identified as a significant fail-
ure mode shortly after their widespread release. Early specialised investigations into
LLM behavior estimated that between 30% and 69% of references generated in bio-
medical and scientific contexts were entirely fabricated (Athaluri et al., 2023; Walters
& Wilder, 2023). Recent systematic audits demonstrate that this has since escalated
into a rapidly escalating global crisis: an analysis of 2.5 million biomedical papers
revealed a more than 12-fold increase in the rate of fabricated references between
2023 and early 2026 (Topaz et al., 2026).
It is important to emphasize that modern Large Language Models (LLMs) are
fundamentally not information-processing tools, but language-processing systems.
They mimic the linguistic patterns and structures of texts—rather than the substan-
tive factual content—found in their training datasets (Walters & Wilder, 2023). Con-
sequently, all such models are prone to “hallucinations” — generating factually in-
correct but plausible-sounding responses. Current efforts to mitigate these inherent
flaws involve the integration of external knowledge retrieval (RAG) and real-time
web-search capabilities to ground outputs in verifiable data. However, as the core
generative engine remains probabilistic, the challenge of ensuring factual accuracy
remains a persistent concern.
This article analyses a real precedent encountered by our journal’s editorial board.
A year ago, the editorial board received a review article manuscript from a group
of authors, dedicated to ecological studies of one of insect groups in North Africa.
The processing of this manuscript revealed all the chronic diseases of modern scien-
tific periodicals.
A severe shortage of reviewers. We faced mass refusals from specialists to review
due to excessive workload.
Superficial peer review. The manuscript eventually underwent double-blind
peer review. Both reviewers provided positive feedback, having spent an inadequate-
ly long time analysing the text, but, as it turned out later, performed the check pure-
ly nominally.
Editorial routine and pressure. Having received positive reviews, the article was
preliminarily accepted for publication. A lengthy and meticulous proofreading stage
began: text editing, checking for compliance with British English, and bringing the
manuscript in line with the journal’s requirements. This process was accompanied by
constant pressure from the authors, who regularly bombarded the editorial office
with emails demanding to expedite the publication: “What is the status of the article
now? When will it be published online?”
The turning point that collapsed this house of cards was the final technical check
of the reference list. The production editor’s attention was drawn to the lack of work-
ing DOIs in a significant portion of the modern links. A selective manual check of
the sources in journal archives yielded a shocking result: a substantial part of the
cited publications did not physically exist. They turned out to be the product of arti-
ficial intelligence hallucinations (most likely early versions of GPT, used without
search plugins).
Phantom references and the peer-review crisis: how artificial intelligence tests
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
303
A detailed analysis of the fabricated reference list allowed us to identify three
main mechanisms of “hallucinations” of early versions of generative language mod-
els, which the authors uncritically integrated into their text and which every editor
should know.
Anachronisms of volumes and issues. AI compiles a real journal name
with random volume numbers and years. For example, the list featured the 12th vol-
ume of the journal Entomological Research for 2020, whereas, in reality, the 50th
volume of this publication was issued in 2020. The model is unable to verify the ac-
tual publishing history.
“Frankenste inisat ion” of sources. The most insidious type of forgery. The
algorithm takes a real journal, correct year and volume, real page ranges, and even
real specialist authors, but completely invents the article title so that it perfectly fits
the narrative of the manuscript. In our case, on the pages of the journal Agronomy for
Sustainable Development, which supposedly contained a review of Orthoptera ecol-
ogy, an article by completely different authors about weed control in legume crops
was actually published.
Profess ional and spat ia l displacements. The model does not distin-
guish the narrow specialisation of scientists. In a falsified reference to Biodiversity
Journal about Moroccan insects, the co-author turned out to be a real and very fa-
mous Moroccan professor-agronomist whose specialisation is herbology (weed sci-
ence), not entomology. The specified pages of the journal actually contained works
on the fauna of Italy and India.
The scale of the problem is corroborated by independent audits. Research
into biomedical literature over the last three years indicates very many of refer-
ences generated by large language models (LLMs) in biomedical contexts are
entirely fabricated (Topaz et al., 2026). This is not merely a technical glitch but a
systemic phenomenon, as these references are often impeccably formatted and
attributed to real researchers, making them virtually “invisible” to traditional
peer review.
C onclusions and warnings for the sc ient i f ic community. Effective-
ly, we are entering a stage where one of the primary challenges in science is no longer
accessing information, but verifying whether this information actually exists.
Our position is categorical: if authors falsify sources so carelessly or deliberately,
there is no reason to trust their own results and primary data. They are just as likely
to be fabricated. This case requires wide publicity within ethical limits. The discovery
of a completely fabricated reference list inevitably casts doubt on the reliability of any
results presented in the article. The manuscript was immediately rejected already at
the typesetting preparation phase, as publishing such material would deal an irrepa-
rable reputational blow to the journal, despite prior approval by reviewers. This case
requires publicity within ethical boundaries (without revealing the names of the au-
thors of the unpublished manuscript), as it exposes a global problem.
The traditional blind peer-review system is built on the presumption of the au-
thor’s integrity. Today, this presumption is dead. Reviewers are not prepared to spend
hours fact-checking every reference, which academic fraudsters actively exploit. We
V. O. Kharchenko, V. O. Korneyev & N. S. Filimonova
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
304
are entering an era of new challenges, where the uncontrolled use of large language
models can destroy the foundation of scientific trust.
The experience of large-scale audits (e.g., scanning 2.5 million papers in Pu-
bMed Central) proves that manual bibliography verification is no longer sufficient. A
viable solution lies in implementing automated reference validation systems directly
into the editorial workflow, as proposed in our updated editorial control algorithm.
Addressing our fellow editors of academic journals, we urge the editorial boards
of all academic journals to implement new control protocols:
S elec t ive checks: mandatory manual verification of 3–5 sources from differ-
ent parts of the reference list (especially those lacking a DOI). Traditional peer re-
view, based on the presumption of authors’ academic integrity, has proven powerless
against generative AI. Reviewers, focusing on the logic of the text, no longer check
the physical existence of sources, trusting properly formatted reference lists.
Mandator y D OI va l idat ion pr ior to peer review. Technical editors
must carry out mandatory automated or many validation ofseveral sources from dif-
ferent parts of the reference list (especially those lacking a DOI). Manuscripts with-
out active digital object identifiers for modern sources must be returned to the au-
thors even before peer review begins.
Declarat ion of AI. The author guidelines must enshrine a strict requirement:
any use of language models (even for translation or formatting) must be detailed in
the Acknowledgements or Methods section, specifying exactly how the generated
information was verified. The absence of such a declaration upon the discovery of AI
traces should be grounds for a lifetime ban of the authors from the publication.
Br ief ing reviewers. Update reviewer forms by adding a point on the manda-
tory selective verification of 3–5 sources from the reference list regarding their phys-
ical existence.
The scientific community must acknowledge: artificial intelligence tools are an
irreversible reality of science, they are already here, and ignoring them is impossible.
But without the introduction of strict verification barriers at the editorial level, sci-
entific literature risks drowning in a flood of plausible but completely empty simula-
cra. It depends solely on our vigilance whether a scientific article will remain the
gold standard of verified knowledge or turn into a generated illusion.
A few words about the “nuts and bolts”. Artificial intelligence does
not feel an “unwillingness”to work and is not “lazy” due to an influx of millions
of users. Models do not have consciousness to choose the path of least resistance
out of fatigue, at the core of this lies a fundamental mathematical problem of the
transformer architecture — Maximum Likelihood Estimation. LLMs (large lan-
guage models) are trained to predict the next word so that the text looks as nat-
ural as possible. It is mathematically “easier”(i. e., requires less computational
effort from the algorithm) for them to generate a perfectly smooth, syntactically
correct, but fabricated article title than to search within their weights for an ex-
act, specific, and rare fact. Furthermore, the so-called “laziness” (when a model
gives short, superficial answers) is often a consequence of alignment procedures
(RLHF — Reinforcement Learning from Human Feedback), where developers
artificially restrict models so they do not generate unsafe content or to save com-
putational resources (tokens) on company servers.
Phantom references and the peer-review crisis: how artificial intelligence tests
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
305
However, over the past year, the landscape of AI tools for scientists has changed
dramatically. Modern researchers rarely use “bare” ChatGPT for literature searches,
as the aforementioned authors did a year ago. Here is the current arsenal now being
used in the academic environment, which brings both benefits and new threats:
Basic universal LLMs (text and code generators):
ChatGPT (OpenAI, GPT-4o models): Has web access and plugins for
data analysis. Used for drafting, structuring articles, and writing scripts (Python/R)
for processing statistical data.
Gemini (Google , Pro/Advanced models): Integrated with the Google
ecosystem (including Google Scholar via an extension). Capable of processing vast
arrays of text. Used for analysing large PDF files. Google’s separate NotebookLM tool
is now massively used by scientists to create personal knowledge bases from upload-
ed articles.
Claude (S onnet 4 .6*): Has a massive context window, allowing authors to
upload a dozen articles and request a compiled review.
Specia l ised sc ient i f ic AI search engines (RAG systems):
Perplexity AI: The most popular search tool at the moment. It does not sim-
ply generate text but searches for real sources online and compiles an answer with
direct links.
E l ic it , C onsensus , S ciSpace: Highly specialised tools trained exclusively
on databases of scientific articles (Semantic Scholar, etc.). They analyse PDFs, “ex-
tract” methodology, results, and create summary tables.
Evolut ion of i l lus ions : f rom pr imit ive errors to per fec t fa ls i f ica-
t ions. It should be understood that the manuscript, which became the subject of our
analysis, was created about a year ago. At that time, the authors likely used basic
versions of language models without direct internet access, which led to the appear-
ance of obvious “hallucination” markers — non-existent volumes and pages. Howev-
er, over the past year, tectonic shifts have occurred in the field of artificial intelli-
gence.
Today, the academic world is massively armed with a new generation of AI tools.
Universal models (ChatGPT-4o, Gemini Advanced, Claude 3.5) with huge context
windows capable of analysing dozens of uploaded articles simultaneously have been
joined by specialised scientific search engines and analysers (Perplexity, Elicit, Con-
sensus, SciSpace). They use RAG (Retrieval-Augmented Generation) technology,
relying on real databases of scientific publications.
Expectedly, the technological leap did not diminish the threat to academic in-
tegrity, but merely made it more insidious. A new systemic problem has emerged: a
fundamental discrepancy between how algorithms work and what users expect from
them. Large language models by their nature are neither knowledge bases nor search
engines — they are probabilistic text generators. Their mathematical goal is to create
a syntactically perfect and statistically plausible sequence of words. When a model
encounters a complex, highly specialised query (for example, searching for empirical
data on local fauna), searching for the exact fact within the neural network’s weights
requires considerable “effort”. It is mathematically easier for the algorithm to take the
path of least resistance: compile an averaged, maximally plausible text that looks like
V. O. Kharchenko, V. O. Korneyev & N. S. Filimonova
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
306
science, sounds like science, but contains no actual facts. This effect is amplified by
safety and optimisation settings from the developers themselves (RLHF), which
force models to provide generalised, superficial answers instead of deep analysis.
As a result, we get a perfect trap for dishonest or lazy authors: AI generates flawless-
ly written English text with perfect structure and actually existing references (thanks to
new search tools), but the conclusions, correlations, or synthesis of ideas themselves are
completely empty, “averaged” hallucinations of a higher order. Detecting such a generat-
ed review or discussion is much more difficult: the bibliography will be genuine, but the
link between the cited source and the author’s thesis may be completely distorted by the
algorithm, which simply tailored the text to the required narrative. Accordingly, the bur-
den of verification again falls on the reviewer, who will now have to verify not only the
existence of the article but also whether it actually states what the author claims. Modern
AI can perfectly imitate form, so the protection of scientific knowledge must be based
exclusively on strict control of content and primary data.
New a lgor ithm for editor ia l ver i f icat ion: how to recognise
hidden generat ion. Since modern RAG systems (AI-based search tools) have
made it almost impossible to detect fake sources by external signs, editorial
boards and reviewers need to shift their focus. Falsifications are moving from the
level of bibliography to the level of meanings and primary data. We propose an-
other algorithm for verifying manuscripts to detect the uncritical or fraudulent
use of the latest AI models:
1. Search for “semantic hallucinations” and the effect of “perfect emptiness”
Modern LLMs generate syntactically flawless but conceptually “sterile” texts. A re-
viewer should look out for the following markers:
Absence of sc ient i f ic conf l ic t: AI algorithmically avoids cognitive disso-
nance and “inconsistencies” in hypotheses. If a review article or discussion section
perfectly reconciles all sources and contains no analysis of contradictory or anoma-
lous results, this is a marker of generated averaging.
Gap between c itat ion and conclusion: The source is real, but it does not
support the thesis put forward by the author. AI often takes a real article but “in-
vents” a conclusion needed for the smoothness of the current narrative. The review-
er must selectively verify not the fact of the article’s existence, but the essence of the
conclusion drawn from it.
Methodologica l s ter i l i ty: Generated descriptions of materials and meth-
ods often look like idealised protocols from textbooks, devoid of the specific rough-
ness of real field or laboratory work.
2. Verification of primary materials (Technical barrier). The most vulnerable
point of artificial intelligence is its inability to generate a flawlessly consistent array
of raw, empirical data with real physical parameters. Therefore, editorial boards must
introduce a presumption of openness for primary materials. A manuscript should
not be admitted to the peer-review stage without fulfilling the following conditions:
Deposit ion of raw data: All measurement tables, feature matrices, and sta-
tistical samples must be uploaded to open repositories (e. g., Zenodo or Dryad) by
the time the article is submitted. Journals must abandon the practice of “data avail-
able upon request”.
Dig ita l and physica l footpr int: For biological research, the provision of
Phantom references and the peer-review crisis: how artificial intelligence tests
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
307
accurate geographical metadata is required. For instance, collection coordinates
must be provided in a strictly standardised format, such as [0.717° S, 77.567°W],
which complicates the automatic generation of random numbers.
Inventor y of mater ia l: Every mentioned specimen or sample must contain
a link to a real voucher number in a recognised depository (while paying attention
to minor formal details that certify the author’s manual work, for example, that the
depository abbreviation ends with a full stop.). The absence of collection inventory
numbers is grounds for immediate rejection.
3. Requirements for analytical tools if authors claim to have conducted complex
statistical or bioinformatic analysis, the requirement for them must be categorical:
Open code: Provision of original scripts (R, Python) and program execution
logs. AI perfectly writes code on demand, but often this code generates perfect
graphs on fake datasets. Providing logs allows verifying whether the code was actu-
ally applied to the declared raw data.
Conclusion
It is time to admit: the era when it was enough to send a well-formatted text (PDF/
Word) for publication has ended. The response to the challenge of generative models
must be a transition to the concept of the Open Science Framework, where the text
of the article is merely an accompanying note to a verified array of primary data. If
an author cannot support their results with raw data, vouchers, and calculation algo-
rithms, no perfectly written English text has the right to be called a scientific publi-
cation.
Declarat ion on the use of ar t i f ic ia l intel l igence and Acknowl-
edgements. The authors officially declare that this editorial text (including the for-
mulation of new requirements for authors and the conceptualisation of the problem)
was generated, structured, and stylistically polished in direct dialogue with the large
language model Gemini Pro. We make this statement not for the sake of an ironic
metaphor, but as a demonstration of the profound and highly uncomfortable crisis
of our time. We have just rejected an article by dishonest authors for delegating ana-
lytical work to an algorithm, and we ourselves — simultaneously delegated the crea-
tion of a text about the inadmissibility of such actions to an algorithm.
This fact should frighten the scientific community far more than a fabricated
reference list. The fact that a machine is capable of flawlessly, with the required level
of academic anger, structural logic, and imitation of human principledness, writing a
manifesto against machine falsifications proves: the boundary between authorial text
and machine generation is finally erased. We demand transparency from authors, so
we start with ourselves. Responsibility for the ideas, implemented rules, and rejection
of the aforementioned manuscript lies wholly and entirely with the human authors.
However, we acknowledge: the toolkit for expressing academic thought is no
longer exclusively human. If we do not start strictly controlling primary data today,
tomorrow it will be technically impossible to distinguish real science from generated
imitation.
The authors express their sincere (and not without a touch of academic irony)
V. O. Kharchenko, V. O. Korneyev & N. S. Filimonova
ISSN 2707-725X. Zoodiversity. 2026. Vol. 60, No. 3
308
gratitude to the large language model Gemini Pro for the ruthless dissection of its
own digital “relatives” algorithms, revealing the anatomy of machine hallucinations,
and assisting in structuring this text. The fact that an article about the existential
threat of uncontrolled artificial intelligence was analysed and written in dialogue
with artificial intelligence is a classic embodiment of the mythological Ouroboros,
the serpent eating its own tail.
However, this paradox brilliantly proves our main postulate: AI is neither an
independent evil nor a panacea. It is merely a hyper-powerful “optical instrument”.
In the hands of a dishonest author, it generates convincing mirages, but under the
strict control of a critical human mind — it becomes a microscope capable of de-
bunking these mirages. The responsibility for every word, conclusion, and rule in-
troduced in this article, as is appropriate in genuine science, is borne exclusively by
the human authors.
REFERENCES
Athaluri, S. A., Manthena, S. V., Kesapragada, V. S. R. K. M., Yarlagadda, V. L., Dave, T. & Duddumpu-
di, R. T. S. 2023. Exploring the boundaries of reality: investigating the phenomenon of ar-
tificial intelligence hallucination in scientific writing through ChatGPT references. Cureus,
15, e37432, 1–5.
https://doi.org/10.7759/cureus.37432
Topaz, M., Roguin, N., Gupta, P., Zhang, Z. & Peltonen, L.-M. 2026. Fabricated citations: an audit
across 2.5 million biomedical papers. The Lancet, 407, 1779–1780. https://doi.org/10.1016/
S0140-6736(26)00603-3
Walters, W.H. & Wilder, E. I. 2023. Fabrication and errors in the bibliographic citations gener-
ated by ChatGPT. Scientific Reports, 13, 14045, 1–9.
Received 14 May 2026
Accepted 30 June 2026
|
| id | oai:ojs.akademperiodyka.org.ua:article-971 |
| institution | Zoodiversity |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2026-06-30T01:00:34Z |
| publishDate | 2026 |
| publisher | Publishing House "Akademperiodyka" of the National Academy of Sciences of Ukraine |
| record_format | ojs |
| resource_txt_mv | ojsakademperiodykaorgua/c4/2a3f6436945532eaa71b21fc7edf7ec4.pdf |
| spelling | oai:ojs.akademperiodyka.org.ua:article-9712026-06-29T16:20:36Z Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals Kharchenko, V. O. Korneyev, V. O. Filimonova, N. S. academic integrity LLM generative hallucinations open science data deposition research falsification The critical vulnerability of the traditional blind peer-review system to the challenges posed by the rapid development of generative tools is examined. Based on a real-world precedent involving the discovery of a completely fabricated bibliography in a submitted manuscript, we analyse the basic mechanisms behind the creation of false references, such as anachronisms, “Frankensteinisation”, and professional biases. The study demonstrates the evolution of the threat: a transition from the obvious errors of early algorithms to the deep “semantic hallucinations” of modern RAG-based search engines, which are capable of generating perfectly formatted yet conceptually empty texts derived from real databases. To protect the publication process, an updated algorithm for editorial control is proposed, requiring the mandatory validation of digital object identifiers (DOIs) and a clear declaration of the algorithms utilised by the authors. The main conclusion emphasises the necessary and unalterable transition to the Open Science Framework paradigm, where textual material is viewed merely as an accompanying document to a verified array of primary datasets, open-source code, and deposited collection specimens. Publishing House "Akademperiodyka" of the National Academy of Sciences of Ukraine 2026-05-14 Article Article application/pdf https://ojs.akademperiodyka.org.ua/index.php/Zoodiversity/article/view/971 10.15407/zoo2026.03.301 Zoodiversity; Vol. 60 No. 3 (2026): Zoodiversity Zoodiversity (Vestnik Zoologii); Том 60 № 3 (2026): Zoodiversity 2707-7268 2707-725X 10.15407/zoo2026.03 en https://ojs.akademperiodyka.org.ua/index.php/Zoodiversity/article/view/971/400 Copyright (c) 2026 Valery Korneyev |
| spellingShingle | Kharchenko, V. O. Korneyev, V. O. Filimonova, N. S. Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title | Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title_full | Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title_fullStr | Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title_full_unstemmed | Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title_short | Phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| title_sort | phantom references and the peer-review crisis: how artificial intelligence tests the resilience of scientific periodicals |
| topic_facet | academic integrity LLM generative hallucinations open science data deposition research falsification |
| url | https://ojs.akademperiodyka.org.ua/index.php/Zoodiversity/article/view/971 |
| work_keys_str_mv | AT kharchenkovo phantomreferencesandthepeerreviewcrisishowartificialintelligenceteststheresilienceofscientificperiodicals AT korneyevvo phantomreferencesandthepeerreviewcrisishowartificialintelligenceteststheresilienceofscientificperiodicals AT filimonovans phantomreferencesandthepeerreviewcrisishowartificialintelligenceteststheresilienceofscientificperiodicals |