Introduction
Academic publishing is facing a flood of low-quality, machine-generated material often described as AI slop. Reports indicate that by 2025, more than half of published academic articles showed signs of large language model involvement, up from roughly one in eight just two years earlier. In some journals, over one in ten peer reviews appeared to be written entirely by AI. With disclosure rates lagging far behind actual usage, the true scale of the problem is likely much larger.
These numbers raise a pressing question: how difficult is it to produce a convincing but entirely fabricated research paper, and do publishers have adequate defenses to keep it out of the scientific record?
What Happened
Writer Benjamin Skuse decided to find out by running an experiment with IEEE Access, a multidisciplinary open-access journal. Working with the journal's managing editor, Jenny Mahoney, he attempted to submit two AI-produced papers neither of which described real, conducted research.
For the first attempt, Skuse used only free versions of ChatGPT and Gemini. He chose optical solitons, a topic from his own doctoral studies years earlier, and prompted the models to generate hypotheses, text, figures, citations, and formatting with minimal human intervention. When one chatbot refused to produce synthetic data, he switched to the other and rephrased his requests. After roughly two dozen prompts, he had a manuscript that looked polished and scientifically plausible.
When submitted, the paper triggered integrity checks. Automated tools and human editors identified numerous fabricated or unverifiable references, including one attributed to John Smith. An expert in optical solitons reviewed the manuscript and found the writing professional but flagged dense mathematics, excessive acronyms, and a few confident physics claims that were at best debatable. The paper was not accepted.
For the second round, Skuse used FARS, an autonomous research system developed by Singapore-based Analemma. With a single prompt, FARS produced a detailed research proposal in about 21 minutes, ran a computational experiment overnight, and drafted a submission-ready article in two hours. Skuse removed the system's self-disclosure warning, added his own name, and sent it to IEEE Access.
This manuscript contained no fake references, but it relied too heavily on non-peer-reviewed preprints. A human editor also spotted a figure labeled with the nonsense term "lhewl" and another image showing an AI brain with random dots. Like the first paper, it was rejected.
Why This Matters
The experiment illustrates a broader crisis in scholarly communication. Science operates under a publish-or-perish culture that rewards volume, a dynamic long exploited by paper mills. Artificial intelligence now acts as an accelerant, making it easier to generate manuscripts at scale. Peer reviewers, who perform this essential gatekeeping work as volunteers, are already overwhelmed. IEEE Access alone receives roughly 60,000 submissions annually, and the number of reviewer invitations per article has climbed by half since 2018.
As automated research agents grow more capable, the line between legitimate science and AI fabrication is blurring. Systems such as FARS, AutoResearchClaw, The AIScientist-v2, and others can conduct literature reviews, design experiments, and write papers with almost no human input. Some are even connecting to autonomous laboratories that produce real-world data.
Key Takeaways
- Free, general-purpose AI tools can generate a superficially credible scientific paper from research that never happened.
- Publisher integrity checks can catch fabricated references, but more advanced AI systems may produce fewer obvious errors.
- Human editors and peer reviewers remain a critical last line of defense, yet they are struggling with rising submission volumes.
- Autonomous research agents can now perform literature searches, experiments, and drafting with minimal human involvement.
- The scientific community faces difficult questions about trust, reproducibility, and what counts as legitimate research in an AI era.
Conclusion
Neither of Skuse's manuscripts was published, and neither underwent genuine peer review. Still, the exercise revealed that an amateur with no specialized expertise can create a paper that appears convincing at first glance. As AI tools improve, detecting forgeries may become only part of the challenge. The harder question may be deciding what qualifies as legitimate scientific work when machines can conduct research and write it up with little human oversight.
Publishers, institutions, and researchers are beginning to explore new safeguards, from blockchain-based trust markers to sample databases with unique identifiers. But the central lesson is clear: business as usual is no longer viable. Science must adapt to a world where distinguishing real research from AI-generated fabrication is increasingly difficult.

Discussion
Join the conversation
Thoughtful reactions, questions, and follow-up ideas help shape the next story.