Benjamin Skuse went back to optical solitons, the subject of his PhD nearly two decades ago, to see whether artificial intelligence could make a piece of research that looked real. The science writer holds a PhD in applied mathematics from the University of Edinburgh and an MSc in science communication from the University of the West of England.
His test used only the free versions of ChatGPT and Gemini. Skuse was barred from editing their output or seeking help from specialist AI-prompt engineers, and worked with roughly two dozen prompts. When one tool refused a task, he turned to the other.
Skuse split the manuscript into its familiar components, including the abstract, introduction and methods, then asked the tools to tackle them in sequence. A dozen or so prompts produced a realistic-looking paper. A further dozen or so created and polished its figures and captions, then supplied references and citations.
He then submitted the paper to IEEE Access with the journal’s agreement, putting its automated integrity checks to work on a manuscript generated through free chatbots. The system flagged fabricated and unverifiable references. Jenny Mahoney, the managing editor of IEEE Access, oversaw the journal’s automated checks.
The result showed one clear weakness in the generated paper. Fabricated references are easier to detect than stylistic signs associated with AI.
Skuse also put FARS through the same exercise. The autonomous research system is designed to carry out much of the scientific work itself, and produced a submission-ready article two hours after Skuse clicked “Write paper”. IEEE Access flagged that paper for its heavy reliance on non-peer-reviewed preprints.
The experiment comes as AI’s presence in academic publishing has become far harder to dismiss as an occasional oddity. By 2025, an estimated 57 percent of published academic articles showed detectable evidence of large language model influence, up from 12 percent in 2023. In 2025, over 12 percent of Nature Communications reviews were found to be completely AI-generated.
Disclosure has lagged far behind the apparent scale of use. For every 40 papers with statistical evidence of AI use in 2025, only one formally disclosed it.
Amanda Sulicz, IEEE’s research-integrity manager, put the concern in broader terms: “It's not just the [scientific] record under threat, it's science as a whole. AI slop is completely eroding trust in science and every scientific process that we've worked for centuries to create.”
Skuse’s tests did not establish whether either paper would have survived peer review. They did show that a person using free tools, a familiar subject and a limited run of prompts could produce a manuscript convincing enough to require a journal’s integrity systems to sort the plausible from the verifiable.
Comments (0)