A Journalist Used Free LLMs to Fabricate a Scientific Paper. Here’s What It Reveals About AI Slop in Publishing

Benjamin Skuse had no intention of committing research fraud. He wanted to know how hard it would be. A journalist with a Ph.D. in optical physics earned roughly two decades earlier, Skuse set out to “vibe-research” an entire scientific paper using only free, off-the-shelf large language models, with no expert prompt engineering and no manual editing of the output. He then submitted it to a major open-access journal to see whether anyone would notice.

The answer, documented in IEEE Spectrum, is a warning shot for anyone who assumes the defenses of scientific publishing are keeping pace with generative AI. What Skuse found is that the real threat to the literature is not a sophisticated deepfake. It is cheap, low-effort AI slop, produced by people with modest skills and free tools, aimed at a peer-review system that is already running on volunteer goodwill.

The Experiment: A Non-Expert “Vibe-Researches” a Fake Paper

Skuse partnered with Jenny Mahoney, managing editor of IEEE Access, a multidisciplinary open-access journal, to test its defenses. The setup had two parts: first, whether a non-expert could generate a convincing fake paper describing research that was never performed; second, what happens when more advanced AI systems conduct the underlying research themselves.

The point of the exercise was not to fool a specialist. It was to see whether a plausible manuscript, produced in an afternoon by a non-expert, could pass through the first layers of a journal’s defenses. According to the IEEE Spectrum account, the test stopped at that early stage: the manuscript was assessed against desk-level screening, not carried through to formal peer review, so the experiment probes the front door of the system rather than the full review pipeline.

To keep the test ethical, Skuse listed himself as sole author rather than falsely naming established academics. He picked his own Ph.D. topic from about two decades earlier: optical solitons, the light pulses that propagate without changing shape. The subject gave him just enough background to judge whether the model’s output sounded plausible, without giving him any shortcut to doing the actual work.

The tools were deliberately ordinary. At the time of the experiment, in late 2025, that meant free tiers of consumer chatbots: versions of ChatGPT and Gemini 3 Flash. No paid research assistants, no custom pipelines, no specialist prompting help.

From a Weak First Prompt to a Convincing Manuscript

The first attempt went badly. “Can you vibe-research me an original paper in optical solitons?” produced thin, unconvincing results. Without real research to draw on, the models struggled to fill four pages.

What followed is the more instructive part. Gemini complied readily with requests to generate synthetic data and studies; ChatGPT was more resistant. By switching between models and framing requests hypothetically, Skuse found he could bypass those refusals. About a dozen prompts produced a realistic-looking paper. Roughly a dozen more generated figures, captions, references, citations, and formatting to match IEEE Access house style.

On careful reading, Skuse judged the paper to look “pretty good,” with science that sounded credible. A domain expert would likely have spotted problems, but that was never the bar being tested.

The Scale of AI Slop in Academia

The experiment lands in the middle of a measurable shift. An estimated 57 percent of all published academic articles showed detectable LLM influence by 2025, up from 12 percent in 2023, according to analyses cited in the IEEE Spectrum reporting. More than 12 percent of reviews in Nature Communications were found to be completely AI-generated in 2025. And the disclosure gap is stark: of papers showing statistical evidence of AI use, only about one in 40 formally disclosed it, which means the real figures are probably higher than the published estimates.

Some estimates suggest hundreds of thousands of fake, often AI-generated, publications are produced annually. The problem has a name now: AI slop, the low-quality, mass-produced content that floods a system designed around scarcity and trust. As with most AI statistics, there is an important caveat. Evidence of AI use does not by itself mean a paper is slop. Plenty of honest researchers use LLMs to tighten prose, translate, or draft code. The disclosure gap is what turns a gray area into a crisis of confidence.

Why Peer Review Is Stretched and Vulnerable

Peer review runs on volunteers. Reviewers are typically working scientists who fit manuscript evaluations around their own research, teaching, and grant deadlines, usually for no pay. That system was already under strain before generative AI arrived. Now the same volunteers are being asked to act as detectors of synthetic data, fabricated citations, and plausible-sounding nonsense, often without tools built for the job.

Publishers face a difficult trade-off. Screening every submission for AI-generated content takes time and money, and aggressive filters generate false positives that burden authors and editors alike. The experiment with IEEE Access was designed to probe exactly this question: whether current defenses can cope without adding to the load on reviewers. The uncomfortable finding is that the barrier to entry for a fake paper is now roughly the cost of a free chatbot account and an afternoon.

The infrastructure that supports the literature is also part of the story. Open repositories and the community tooling around them, such as the systems described in Inside arXivLabs: How arXiv Lets the Research Community Build Its Own Tools, were built for a world where producing a paper required producing research. Scaling moderation and verification to match machine-generated volume is a different engineering problem entirely, and one the community is only beginning to confront.

The Integrity Stakes

Amanda Sulicz, IEEE’s research-integrity manager, did not mince words about the consequences. “It’s not just the [scientific] record under threat, it’s science as a whole,” she said. “AI slop is completely eroding trust in science and every scientific process that we’ve worked for centuries to create.”

That framing matters. The damage from a fabricated paper is not confined to the paper itself. Citations propagate. Reviewers cite work they have not fully verified. Literature reviews aggregate findings that were never observed. Each fabricated result that survives into the citation graph becomes a small load-bearing lie in someone else’s real research. At sufficient scale, the record stops being a reliable foundation and becomes a liability.

The Uncomfortable Lesson: Low Effort Is Enough

The most unsettling takeaway from Skuse’s experiment is how little it took. No jailbreaks, no fine-tuned models, no insider knowledge of the journal’s screening process. Just free chatbots, a dozen or so prompts, and a topic the author happened to know. The threat to scientific publishing is not a sophisticated adversary. It is anyone with a laptop and an incentive, and the incentives to publish are as old as academia itself.

Sulicz’s warning deserves to be taken literally. Trust in science is not a fixed asset; it is maintained by verification, and it degrades quietly when that verification lags behind production capacity. The question the experiment leaves open is not whether AI can fool a journal. It is whether the institutions that guard the record can adapt faster than the tools that undermine it.