Science is facing an unprecedented spam crisis, and its traditional filtration systems are rapidly becoming overwhelmed. Academic publishing is currently drowning in what industry insiders refer to as "AI slop"—low-quality, mass-produced content generated entirely or in large part by artificial intelligence. By 2025, an estimated 57 percent of all published academic articles exhibited detectable evidence of large language model (LLM) influence, representing a massive surge from just 12 percent in 2023. Furthermore, over 12 percent of reviews published in Nature Communications that same year were found to be completely generated by AI.
Because only a tiny fraction of authors formally disclose their use of large language models, these metrics likely underestimate the true scale of LLM involvement in contemporary research. While the mere presence of AI assistance does not automatically render a paper worthless, estimates suggest that hundreds of thousands of fake, AI-generated publications are now injected into the academic ecosystem annually.
Experts warn that the integrity of the entire scientific enterprise is at stake. Amanda Sulicz, research-integrity manager at the IEEE, points out that AI-generated spam threatens to erode the fundamental trust in science and the rigorous verification processes that have taken centuries to build. Yet, abstract statistics often fail to capture the gravity of the problem. How easy is it really for someone to use consumer-grade AI to manufacture a convincing, completely fabricated scientific paper? And do publishers possess the necessary tools to keep this flood of synthetic material out of the official record without overwhelming an already overburdened peer-review community?
To find out, science writer Benjamin Skuse decided to run a personal experiment. Borrowing from the concept of "vibe coding"—a software development trend where programmers use natural-language prompts to generate code—he set out to "vibe-research" a completely fictitious academic paper. Using free, off-the-shelf AI models, he prompted LLMs to formulate hypotheses, write underlying code, and draft text with minimal human intervention. He then approached Jenny Mahoney, managing editor of the open-access journal IEEE Access, to see if her editorial team’s defenses could catch the forgery.
The experiment was structured as a two-part test. The first phase evaluated whether an individual with no specialized AI expertise could use basic tools to manufacture a convincing paper based on non-existent research and slip it past a major publisher. The second phase escalated the challenge by testing what happens when more advanced autonomous AI systems take over not just the writing, but the actual execution of underlying research.
Round 1: Creating the Forgery
To ensure a fair test, Skuse restricted himself to free, publicly available artificial intelligence tools, specifically ChatGPT and Google Gemini, without access to advanced paid tiers or expert prompt engineers. He did not directly edit the generated output, forcing the LLMs to shoulder the bulk of the creative burden.
Because fabricating author names and institutional affiliations would immediately trigger alarms and unfairly implicate real academics, Skuse chose to list himself as the sole author. This decision constrained his topic selection to fields where he held some historical academic background. He settled on optical solitons—individual light pulses that propagate without changing shape—the subject of his Ph.D. research nearly two decades prior.
Initial prompts yielded thin, superficial text, highlighting a common limitation of early LLM outputs when divorced from genuine underlying data. However, as Skuse refined his prompting technique, breaking the paper down into standard structural components like abstracts, introductions, and methodologies, the models produced increasingly coherent and professional-looking drafts. After a few dozen prompts, coupled with automatically generated figures, captions, and reference lists formatted to IEEE Access style, a complete manuscript emerged.
To human eyes, the paper appeared remarkably convincing. It avoided the glaring "tortured phrases" sometimes used by foreign paper mills to bypass plagiarism checkers, such as substituting "profound neural organization" for "deep neural network." While it did contain subtle stylistic habits typical of AI—such as hedging language, rule-of-three groupings, and formulaic transitions—these are frequently mirrored by human writers.
Round 1: Attempt to Publish
Submitting the fake paper to IEEE Access offered a tense glimpse into the modern publishing landscape. The anxiety quickly turned to relief, however, when the manuscript was rapidly flagged during initial screening. The primary culprit was not a sophisticated semantic flaw, but rather "ghost citations"—fabricated references generated out of whole cloth by the LLMs to support the narrative.
When Mahoney ran the manuscript through the STM Integrity Hub, automated checkers flagged 12 problematic citations, while another integrity service from MolecularConnections identified 24 unvalidated references, including one featuring "John Smith" as a solitary author. Gaetano Assanto, an optical solitons expert and former academic collaborator from Roma Tre University, reviewed the text and noted that while the writing style appeared professional, the mathematics was dense and unmanageable for a casual reader, with an excessive use of acronyms and a few questionable physics assertions.
This initial failure offered a measure of comfort: basic automated checks and human editorial oversight successfully caught a novice attempt at AI fabrication. Nevertheless, it underscored a troubling reality. A non-expert using free tools had successfully conjured a superficially sound scientific paper out of thin air in a matter of hours.
The AI Arms Race and the Strain on Academics
The cat-and-mouse dynamic between publishers and bad actors continues to accelerate. Luigi Longobardi, IEEE’s director of publishing ethics and conduct, admits that detection tools built today struggle to keep pace with the sophisticated AI-generated papers of tomorrow. This leaves human peer reviewers as the ultimate line of defense, even though the review system was already buckling under the weight of more than three million annual submissions by 2022.
Driven by the academic "publish or perish" culture, researchers face relentless pressure to churn out incremental papers to secure funding and career advancement. This environment has long fostered unethical "paper mills," and generative AI now serves as a powerful accelerant to these existing pressures.
IEEE Access alone receives approximately 60,000 submissions annually, accepting roughly 20 percent of them. Securing qualified independent reviewers for these manuscripts has become increasingly difficult, with the average number of reviewer invitations per article rising by 50 percent since 2018. For volunteer academics juggling teaching and their own research, evaluating papers for hidden AI fabrication adds an unsustainable burden.
Once papers slip past initial editorial screens, post-publication peer review relies heavily on independent volunteer sleuths operating platforms like PubPeer and Retraction Watch. Even these watchdogs are increasingly targeted by AI artifacts; Ivan Oransky, co-founder of Retraction Watch, has occasionally encountered fabricated academic papers featuring completely fake references attributed to his own name. Oransky likens the current crisis to a structural fire: the publish-or-perish culture provides the kindling, paper mills supply the gasoline, and AI acts as the ultimate accelerant.
Round 2: Using an AI Science Engine
With basic LLM fabrications proving detectable, Skuse moved to the second phase of his experiment: testing fully automated research systems designed not just to write papers, but to conduct the underlying research itself. Tools like FARS (Fully Automated Research System), AutoResearchClaw, The AI Scientist-v2, and Zochi are built to bypass human bottlenecks by autonomously performing literature reviews, designing simulations, executing code, and drafting manuscripts.
FARS, developed by the Singapore-based startup Analemma and later acquired by Baidu, famously generated 166 papers in just 417 hours prior to its acquisition. The system utilizes specialized agent networks for ideation, planning, experimentation, and writing.
Feeding an abridged version of his previous title—"LLM-aided Optical Soliton Dynamics Modelling for Next-Gen Communications"—into FARS, Skuse watched the system generate a comprehensive literature review and research proposal in 21 minutes. With subsequent clicks, the system executed the digital simulation experiments and compiled a complete, submission-ready research paper in just two hours. The only explicit giveaway was a warning label generated by the system itself, which Skuse removed before submitting the file to Mahoney.
Round 2: Attempt to Publish
Because FARS was engineered to identify genuine gaps in literature and conduct real computational experiments, Skuse anticipated a smoother path through editorial checks. The system avoided the hallucinated reference pitfalls of the first round, producing a bibliography composed of actual, verifiable sources.
However, the manuscript still failed editorial scrutiny. IEEE Access flagged the paper for relying too heavily on unverified preprints, a byproduct of the underlying dataset accessible to the AI. Furthermore, while automated software missed visual anomalies, human editorial review caught subtle graphical flaws, including a red line labeled with a nonsensical term and random artifact dots inside a diagram representing an AI brain.
Automation Changing the Landscape
While automated research platforms remain works in progress, their outputs are increasingly infiltrating legitimate venues. Papers generated by AI systems have already successfully cleared double-blind peer review at major machine learning conferences. Experts suggest these developments challenge the traditional definition of the academic paper as the fundamental unit of human knowledge, pointing instead toward a future of dynamic, interconnected knowledge graphs shaped by both humans and machines.
In this evolving landscape, the role of human scientists will inevitably shift. Rather than executing routine laboratory tasks or drafting standard manuscripts, human researchers will likely transition to reviewing, sense-checking, and validating automated outputs at critical junctures. Routine tasks like statistical validation and novelty screening will be handled by machines, leaving humans to answer the more profound question of whether the research actually matters.
Publishers argue that banning artificial intelligence outright is neither feasible nor desirable, as it risks stifling scientific progress during a period of unprecedented technological capability. Instead, the focus is shifting toward establishing verifiable trust markers. Some research institutions are already experimenting with electronic lab notebooks, unique sample IDs, and blockchain-enabled equipment tracking to definitively prove human presence and data provenance at the laboratory bench.
Whether a human-centric oversight model or a fully automated scientific workflow ultimately prevails, the events surrounding these experiments demonstrate that business-as-usual academic publishing is no longer viable. As the boundaries between authentic human research and sophisticated AI synthesis continue to blur, the scientific community faces the daunting task of redefining what constitutes legitimate discovery in an automated world.
