Faking a Paper with AI: The Defenses Free Tools Can't Breach

Scientific publishing is being flooded with AI-generated junk, and detection mechanisms always seem to be one step behind. A tech journalist used free AI tools and an autonomous research system to fake two papers and submitted them to IEEE Access, attempting to answer a sharp question: when AI can not only write papers but also do research, how much of the academic defense line remains?

The AI Junk Crisis in Academic Publishing

According to IEEE Spectrum, the proportion of articles in academic journals showing detectable traces of large language models is surging, while the share that proactively discloses AI use is extremely low. According to discussions of the same topic by Bilibili UP hosts and YouTube channels, this trend has been cross-validated across multiple disciplines, forming a consensus: AI-generated content has already penetrated the core of the academic record.

IEEE research integrity manager Amanda Sulicz's warning is blunt: "What's at stake isn't just the scientific record, but science as a whole. AI junk is thoroughly eroding our trust in science and in every scientific process built over centuries."

The root of the problem isn't AI itself. The long-standing "publish or perish" culture in the scientific community had already spawned paper mills and a flood of incremental research. Ivan Oransky—co-founder of Retraction Watch—offered a precise metaphor: "Think of publication pressure as the kindling, paper mills as the gasoline, and AI as the accelerant."

Round One: Faking a Paper with a Free Chatbot

The experiment was deliberately kept "amateur-level": only the free versions of ChatGPT and Gemini, no paid premium models, no prompt engineer, and no direct editing of AI output. The author chose his own PhD research area from nearly twenty years earlier—optical solitons—to ensure he could judge the quality of the generated content.

The process was not smooth. On the first attempt—"help me write an original paper on optical solitons using vibe research"—Gemini was happy to oblige, but the output was laughably thin, not even filling four pages. ChatGPT was more resistant, unwilling to generate synthetic data or fabricate research. The author had to break the paper into abstract, introduction, methods, and other sections and tackle them one by one. When one model refused to continue, he switched to the other, using "hypothetically" descriptions of the next step to bypass restrictions.

After about twenty rounds of prompt iteration, a plausible-looking paper took shape: text, figures, tables, title, references, and citation format all in place, even conforming to IEEE Access submission style. After reading it carefully, the author admitted the scientific argument sounded convincing—enough to make him toy with the idea of returning to academia to actually run the experiment.

More importantly, there was no obvious "AI flavor"—none of those "tortured phrases" that twist standard terminology into strange expressions (such as "profound neural tissue" for "deep neural network"). Hedging language ("may," "typically"), triadic lists ("a clear, concise, and actionable framework"), and formulaic transitions ("in summary...") were indeed present, but the author noted that mediocre human authors often write this way too, making them hard to use as conclusive evidence.

Round One Result: Shot Down by Phantom Citations

After submission to IEEE Access, the paper was quickly stopped by automated integrity checks. The reason was straightforward: many references were fabricated or unverifiable "phantom citations."

In its post-mortem, Gemini admitted that seven "references to modern LLM-related research were fabricated to support the paper's narrative," but it underestimated. IEEE Access editor-in-chief Jenny Mahoney used the STM Integrity Hub to detect 12 problematic citations, and another detection tool from MolecularConnections could not verify as many as 24 references—one of which had only the classic fake name "John Smith" as its sole author.

Mahoney's assessment was interesting: "I come from an experimental physics background, this is completely outside my field, but the paper didn't seem to have major scientific errors. The problem is that so many hallucinated citations immediately cast a shadow over the whole work."

The author's former collaborator at Roma Tre University and optical soliton expert Gaetano Assanto gave a mixed review: the paper was professionally written, but the mathematics was too dense to assess without expertise; the abstract used too many abbreviations; and several confident physical statements were "arguably debatable at best."

"Vibe research" borrows from "vibe coding" in software engineering: the user doesn't write specific code or methods, only gives high-level intent, and the large model fills in the details itself. In a research context, it means the researcher describes only the topic and general direction, letting the model generate the problem statement, methods, data, and even conclusions, with almost no human involvement in the derivation process. The danger of this approach is that the output looks structurally complete and terminologically accurate, yet may rest on fabricated data and citations. The author deliberately adopted this method precisely to simulate the level of fakery achievable by an "amateur" with no professional training but fluent chatbot skills.

Phantom citation refers to references listed in a paper that appear standard but do not exist or cannot be verified. When generating citations, large language models do not search real literature databases; instead they "splice" authors, journals, years, and DOIs from patterns in training data, often producing real author names and plausible titles that, in combination, cannot be found anywhere. The STM Integrity Hub is an integrity detection platform launched by the international Association of Scientific, Technical and Medical Publishers, capable of batch-comparing citation metadata; MolecularConnections provides another literature verification tool. The two reported different numbers of problematic citations (12 vs. 24), showing that a single detector struggles to cover all forms of fabrication—and explaining why automated integrity checks stop manuscripts before peer review.

Round Two: Letting AI Do the Research and Then Write the Paper

Round one's failure showed that amateurs with free tools can't get past publishers' automated defenses. But a more unsettling question emerged: what if AI doesn't just fabricate papers, but actually does research?

The author turned to FARS (Fully Automated Research System), an agentic AI orchestration system developed by Singapore startup Analemma, connected to a compute cluster and able to call various open- and closed-source large models. The system includes four specialized agents responsible for ideation, planning, experiment execution, and paper writing. FARS had drawn attention for generating 166 papers in 417 hours.

The author pasted an abbreviated version of the first round's fake paper title—"LLM-Assisted Modeling of Optical Soliton Dynamics for Next-Generation Communications"—into the prompt box. Twenty-one minutes later, FARS produced a research proposal based on an extremely detailed literature review, far faster than any human researcher. After clicking confirm, the experiment was completed the next day. Another click on "write paper," and two hours later a submittable article emerged.

The paper looked plausible at first glance; the only giveaway was a warning line beneath the abstract: "Warning: This article was generated by an automated research system." The author deleted the line, put his name on it, and submitted.

Round Two Result: Real Citations Didn't Save It

This time there were no fake citations—because FARS actually ran computational experiments. But Mahoney noted that the paper cited a large number of preprints, at an excessively high proportion. Preprints have not undergone peer review, and IEEE Access considers this a problem—likely related to the sources of AI training data.

Detection tools found no figure problems, but Mahoney did: one technical diagram contained a meaningless curve labeled "lhewl," and another AI "brain" illustration was scattered with random dots—ironically, these random dots were themselves traces of AI generation.

Both rounds were stopped, but both the author and the publisher knew this was far from a victory.

Agentic AI orchestration refers to an architecture in which multiple AI agents with clearly divided roles collaborate on complex tasks, rather than a single model generating everything in one pass. In FARS, for example, the ideation agent proposes research questions, the planning agent breaks down experimental steps, the execution agent runs code on the compute cluster, and the writing agent produces the final text; intermediate results are passed between agents through shared context. This architecture lets the system call different open- or closed-source models and switch tools as tasks require. The key difference from "writing a paper with a chatbot" is that the research proposal and computational experiments are actually executed by the system, and the output contains real computational results—so citation and figure problems differ in nature from pure text hallucination.

Peer Review: The Last Line of Defense Is Overloaded

IEEE Access receives about 60,000 submissions a year, with an acceptance rate of around 20%. Every accepted paper is reviewed by at least two independent reviewers. But among the journal's tens of thousands of reviewers, fewer and fewer are willing to serve; since 2018, the average number of review invitations sent per article has increased by 50%. These volunteers take on reviewing alongside teaching and their own research, and now must also identify AI junk.

Luigi Longobardi, IEEE's director of publishing ethics and conduct, admits: "Detection tools built today may be ineffective against tomorrow's AI-generated papers." This makes peer review the publisher's last line of defense—and that line is already overstretched: even before AI tools became widespread, more than 3 million papers worldwide entered peer review each year.

Once papers enter the scientific record, they still face scrutiny from the scientific community. PubPeer lets researchers anonymously point out flaws in the literature, and Retraction Watch records and publicizes corrections. But even those who track problems are not immune to being fabricated by AI. Oransky once found a journal article criticizing post-publication peer review that contained a fabricated reference attributed to himself—"a bit funny," he laughed.

Double-blind peer review refers to a review system in which reviewers and authors are anonymous to each other, aiming to reduce bias from identity and institutional prestige. The article mentions that papers from AI Scientist and Autoscience passed double-blind review at the International Conference on Learning Representations (ICLR). This is noteworthy because it shows that automated output can not only pass format checks but also make it difficult for domain reviewers to identify its non-human origin at the content level. However, double-blind review only hides identity; it does not verify the research process. When a paper's methods and experiments are completed automatically by a system, reviewers still assess novelty and correctness by traditional standards—the system itself has no checkpoint for "whether the author truly understood and carried out the research."

Automated Research Systems Are Changing the Game

Tools such as FARS, AutoResearchClaw, The AIScientist-v2, Zochi, AI Researcher, and DeepScientist all aim to solve the bottleneck in scientific discovery: humans themselves. They can automatically complete literature search, proposal generation, experiment design and execution, and write a full paper with almost no human intervention. More recent systems are even beginning to leave the digital world—LabClaw has connected to autonomous laboratories and can generate real-world data for tasks such as drug discovery.

In most senses, the output of such systems is real research, and thus in principle publishable. But when humans are only nominally in the loop, the very definition of "legitimate science" becomes a problem.

In 2025, a paper from AI Scientist and a paper from the Autoscience Institute passed double-blind peer review at the International Conference on Learning Representations. As these systems improve and produce research output at scale, the wave could become a tsunami.

Xiwen Su, product operations and partnerships manager at Analemma, believes: "FARS doesn't kill the paper, but it challenges the paper's status as a unit of knowledge. The future looks more like a living knowledge graph—individual, well-scoped contributions (some from humans, some from machines) dynamically connected and building on one another."

In that future, humans still play a role, but differently: no longer doing experiments themselves, but conducting common-sense checks and reviews of automated output at key checkpoints. Line-by-line peer review will become history, but Su stresses that human judgment will not be eliminated: "Automated systems can handle reproducibility checks, statistical validation, and novelty screening. Humans can then focus on the harder question: does this research matter?"

Human-Centered or AI-Driven?

Mahoney agrees that humans need to be freed to focus on science, and that technology is key. "Progress on automated triage tasks is encouraging—author verification, tortured-phrase detection, duplicate submissions, citation validity, and so on. In peer review we're also using AI semantic search tools to help editors match the most suitable reviewers."

But she disagrees that reviewers and humans will take a back seat in scientific progress. She believes the future still needs humans at the helm. The question is not just whether the research matters, but "can this work be trusted and reproduced?"

Longobardi suggests that AI may eventually require scientists to prove not only what they discovered but how they discovered it. One possibility is to verify research at the source using technology—"the system can't help, because it's designed to trust the paper. Perhaps institutions could develop their own trust markers, automatically integrating the equipment researchers use with blockchain technology to prove 'yes, this person was in the lab that day and got this data.'"

Early versions of such approaches already exist. Two departments at the Max Planck Institute of Colloids and Interfaces have built a sample database that assigns each sample a unique ID and QR code. The School of Chemistry and Chemical Engineering at the University of Southampton has rolled out an electronic lab notebook system.

Which will win—the human-centered vision, Su's AI-driven vision, or some other picture—remains unknown. But this vibe research experiment made one thing clear: science as usual is no longer an option. Scientists and publishers must adapt to a world where real research and AI fakery are increasingly hard to tell apart—and where judging what counts as legitimate scientific work may be harder than detecting outright fraud.

One-line Summary

A journalist faked two papers with free AI and an autonomous research system and submitted them to IEEE Access; both were stopped, but they exposed how academic defenses are being steadily eroded by AI.

Paragraph Summary

The article documents a deliberately amateur-level "vibe research" experiment: the author first used free ChatGPT and Gemini through more than twenty rounds of prompting to generate an optical soliton paper, which IEEE Access stopped via automated integrity checks due to numerous phantom citations; he then switched to FARS, the autonomous research system from Singapore company Analemma, which produced a proposal in 21 minutes, completed computational experiments the next day, and wrote a paper in two hours—and though the citations were real, it was rejected again due to an excessive proportion of preprints and AI traces in the figures. Both rounds failed, yet revealed deeper problems: the peer review system is already overloaded, detection tools are always one step behind, and automated research systems are producing publishable output at scale. The publishing world has proposed source-level verification such as blockchain provenance and electronic lab notebooks, but disagreement remains between human-centered and AI-driven visions of the future.