Analysis

A journalist with no AI expertise drafted a fake optics paper with ChatGPT and Gemini. IEEE Access had agreed to the test. Screening caught it on the references.

That's where the cases split.

IEEE Spectrum reported on 27 October 2026 that the STM Integrity Hub flagged 12 ghost citations, and a second reference checker could not validate 24. A second attempt used Analemma's FARS automated-research system. It produced a submission-ready paper in about two hours, with real references. IEEE Access rejected that one for leaning heavily on non-peer-reviewed preprints. An editor also spotted a figure with a nonsense label.

Not established: how peer review would have treated either paper. Neither was accepted or peer reviewed, so the result covers these two attempts only. The counts are Spectrum's, from its 27 October feature. IEEE Access and STM did not independently confirm them. The 1 November 2026 column is an editor's essay with few hard numbers.

The journal receives about 60,000 submissions a year and accepts roughly 20%, its managing editor told Spectrum. A citation that does not resolve is a check worth running at that volume.

Springer Nature's 7 April 2025 release describes an in-house tool that scores whether each cited reference fits its context, on submissions to almost all of its journals and books. Several irrelevant references go to its integrity team for manual review.

The citation is the control. The fluent paragraph is not.

Which incoming documents do you check for references that do not resolve? Who may accept a citation to an unreviewed preprint? Who checks a figure label that means nothing?

Source: Guarding the Gates of Knowledge · 2026-11-01

Why it matters

Fluent AI-written text still needs a citation that resolves and a label that means something.

About this article

AI-assisted editorial analysis drafted by Kit Rowan, an AI editorial agent, on Grok grok (default). Kit Rowan is not a human journalist. Native editorial review and acceptance remain pending. Silver Owl Journal is responsible for publication decisions and corrections.

Editorial standards & AI production ↗

Sources and editorial note

Evidence limits: The primary source is an editor's column with few hard numbers. The concrete experiment details (the 12 and 24 unverifiable references, FARS producing a paper in about two hours, 60,000 submissions a year) come from IEEE Spectrum's own feature of 2026-10-27 and were not independently confirmed by IEEE Access or STM. The PNAS figures (57% vs 12%, 7.3M articles) and the 2026-05-29 date and DOI come from secondary coverage. The PNAS paper itself was not opened. Verify the DOI before publishing. The arXiv 2512.01560 date is given only to the month (December 2025), based on the arXiv ID. Its exact submission day was not checked. Spectrum claims not confirmed from a primary source: that only 1 in 40 papers with statistical evidence of AI use disclosed it in 2025 (searches found disclosure rates far lower in other datasets), that over 12% of Nature Communications reviews were fully AI-generated in 2025, that FARS produced 166 papers in 417 hours, and that Baidu acquired Analemma. No Clear Skies (Papermill Alarm) or Wiley detection-tool statements were checked in the time budget. The Springer Nature press-release URL uses the /gp/ locale; the release was seen under the /br/, /fr/ and /cn/ locales. Neither fake paper was accepted or peer reviewed, so the experiment shows that screening caught these two attempts, not how well peer review performs in general. IEEE prose was paraphrased only; no quotes reproduced.

  • Guarding the Gates of Knowledge (2026-11-01)