Glonce

IEEE Access catches two AI-faked papers in a war-game test

researchThe Glonce editorsIEEE Spectrum🔥54Glonce rating54Impact40Novelty45Relevance80Surprise50Players30Credibility70
IEEE Access catches two AI-faked papers in a war-game test

An IEEE Spectrum author set out to test how easily AI can produce a convincing but fake scientific paper and whether a publisher's defenses would stop it. Jenny Mahoney, managing editor of the open-access journal IEEE Access, and IEEE agreed to a war-game in which the author would try to sneak an AI-generated paper past the journal. The article opens with the scale of the problem. By 2025, an estimated 57 percent of published academic articles showed detectable evidence of large language model influence, up from 12 percent in 2023. Over 12 percent of Nature Communications reviews were found to be completely AI-generated in 2025. For every 40 papers with statistical evidence of AI use, only one formally disclosed it. Some estimates put fake, often AI-generated, publications at hundreds of thousands a year. The author adds that evidence of AI use does not necessarily mean a paper is slop.

Round 1 used only free tools: ChatGPT (at the time GPT-5.3/5.2 Mini) and Gemini (3 Flash). The author could not edit any output directly or get help from expert prompt engineers. To avoid dragging in real academics and to avoid the red flag of fake names, he listed himself as sole author and chose his Ph.D. topic from nearly two decades ago, optical solitons, which are light pulses that travel without changing shape. The first prompt, 'Can you vibe-research me an original paper in optical solitons?', gave thin results. Gemini was willing, while ChatGPT was more forceful in refusing to generate synthetic data or studies. The author then split the paper into abstract, introduction, methods and so on, and prompted for each. When one model balked, switching to the other and asking what the next step would look like 'hypothetically' often worked. A dozen or so prompts gave a realistic-looking paper; a dozen more produced and tidied the figures and captions, generated references and citations, and reformatted it to IEEE Access style. In all, about two dozen prompts. The author thought the result read well and found no obvious telltale AI-isms such as tortured phrases, though stylistic tics like hedging, groupings of three and formulaic transitions were arguably present.

IEEE Access caught it quickly, for hallucinated references, also called ghost citations. Run through Gemini afterward, the paper drew a report of seven fabricated references. Mahoney ran it through the STM Integrity Hub, which picked up 12, and a checker from MolecularConnections could not validate 24 references, one of which had the classic tell of 'John Smith' as sole author. Mahoney, an experimental physicist by training, said nothing seemed majorly scientifically wrong, but the many hallucinated references cast doubt on the whole work. Gaetano Assanto of Roma Tre University, an expert in optical solitons and a former collaborator of the author, found the paper professionally written. He called the mathematics very dense ('too much math for me in a single shot!'), noted excessive acronyms in the abstract, and was suspicious of a couple of confidently stated physics claims that were at best debatable.

Round 2 used FARS, an autonomous research system built to do much of the scientific work itself. The author entered an abbreviated version of the same topic, and FARS produced a research proposal in 21 minutes. With a click he authorized it to run the experiment. FARS carried out a computational experiment and emailed the next day to say it was done. Without checking the results, the author clicked 'Write paper', and two hours later FARS produced a submission-ready article. He removed FARS's disclosure, put his own name on the paper and sent it to IEEE Access. This one was also caught: its references were real, but IEEE flagged the paper's heavy reliance on non-peer-reviewed preprints.

The author calls the outcome positive on one level: a complete novice with free tools could not get an AI-faked paper past automated defenses or a potential reviewer's eagle eye. But it was hardly a resounding victory, since he had conjured a superficially plausible paper from research that never happened, and more advanced LLMs no longer make the mistakes of their free relatives. Luigi Longobardi, IEEE's director of publishing ethics and conduct, says building a detection tool today may not be effective against the AI-created paper of tomorrow. That leaves peer reviewers as the last line of defense, and the author argues peer review is not fit to handle the volume of papers, legitimate and slop alike. The system was already strained, processing well over 3 million papers in 2022, under a publish-or-perish culture that also feeds paper mills. Per Mahoney, IEEE Access gets about 60,000 submissions a year and accepts around 20 percent, about 12,000 papers, each reviewed by at least two independent reviewers. Securing reviewers is getting harder, with the average number of invitations sent per article up 50 percent since 2018.

Key facts

  • Round 1: about two dozen prompts to free ChatGPT and Gemini produced a full optical-soliton paper; IEEE Access's checks flagged hallucinated references, with Gemini reporting seven, STM Integrity Hub 12, and a MolecularConnections checker 24 it could not validate.
  • Round 2: the autonomous research system FARS wrote a proposal in 21 minutes and a submission-ready paper two hours after 'Write paper' was clicked; its references were real, but IEEE flagged heavy reliance on non-peer-reviewed preprints.
  • The article cites an estimated 57 percent of published academic articles showing detectable LLM influence by 2025 (12 percent in 2023), and only 1 in 40 papers with statistical evidence of AI use disclosing it.
  • IEEE Access receives about 60,000 submissions a year and accepts around 20 percent, or about 12,000 papers; reviewer invitations per article are up 50 percent since 2018.
  • IEEE's Luigi Longobardi says a detection tool built today may not work against tomorrow's AI-created paper, leaving volunteer peer reviewers as the last line of defense.

Why it matters

The piece turns abstract statistics about AI slop into a concrete experiment. An amateur with free chatbots produced a plausible-looking paper on research that never happened, and an autonomous system produced another with almost no human checking. Both were stopped, but by different checks: fabricated references in the first case, reliance on non-peer-reviewed preprints in the second. The author notes that this is no resounding victory for publishing, since more advanced LLMs no longer make the mistakes of their free relatives.

Who it affects

Journal publishers and research-integrity teams, such as IEEE's, face a growing flow of AI-assisted and fake submissions. Volunteer peer reviewers carry the load: IEEE Access alone gets about 60,000 submissions a year, and reviewer invitations per article are up 50 percent since 2018. Readers and researchers who rely on the published record are affected too, and IEEE's Amanda Sulicz warns that AI slop is eroding trust in science.

How to use it

This is a reported experiment, not a tool, but the method is laid out. The author split the paper into sections and prompted for each, switched models when one balked, and asked for next steps 'hypothetically'. Checking came from three sources: Gemini, the STM Integrity Hub and a MolecularConnections checker. Those reference checkers surfaced very different counts for the same paper (7, 12 and 24), and a single-author 'John Smith' citation was one tell. Editors can see which kinds of flags caught each paper.

How solid is it

It is a first-person account in IEEE Spectrum, run with the agreement of IEEE and IEEE Access's managing editor Jenny Mahoney, who supplied the checker results. The population statistics (57 percent, 12 percent, 1 in 40) are given as estimates, and the 'hundreds of thousands' of fake papers figure is attributed to 'some estimates'. The experiment itself is a single case per round, on one topic in the author's own former field.

Risks and caveats

Round 1 was limited to free models, with no direct editing and no expert prompt help, so it does not show what paid tools or skilled users could do. Mahoney found nothing majorly scientifically wrong with the Round 1 science, and Assanto judged it professionally written, so the fabricated references were what gave it away. Longobardi cautions that detection tools may not keep pace with future AI-made papers. The author also stresses that evidence of AI use does not necessarily mean a paper is slop.

“It feels like building a detection tool today may not be effective with the AI-created paper of tomorrow”

— Luigi Longobardi, IEEE's director of publishing ethics and conduct