vix.ing · top · new · best · stats · spec

Tortured phrases: A dubious writing style emerging in science. Evidence of critical issues affecting established journals

A journal fast-tracked hundreds of papers with garbled tortured phrases, reused images, and abstracts a GPT detector flagged as synthetic.

2021/07/12 by Guillaume Cabanac, Cyril Labbé, Cabanac, Guillaume +3 · 15 voices · 12 citations
#cs.DL #cs.CL #cs.CY #cs.IR

paper · pdf · doi:10.48550/arxiv.2107.06751

Abstract

Probabilistic text generators have been used to produce fake scientific papers for more than a decade. Such nonsensical papers are easily detected by both human and machine. Now more complex AI-powered generation techniques produce texts indistinguishable from that of humans and the generation of scientific texts from a few keywords has been documented. Our study introduces the concept of tortured phrases: unexpected weird phrases in lieu of established ones, such as 'counterfeit consciousness' instead of 'artificial intelligence.' We combed the literature for tortured phrases and study one reputable journal where these concentrated en masse. Hypothesising the use of advanced language models we ran a detector on the abstracts of recent articles of this journal and on several control sets. The pairwise comparisons reveal a concentration of abstracts flagged as 'synthetic' in the journal. We also highlight irregularities in its operation, such as abrupt changes in editorial timelines. We substantiate our call for investigation by analysing several individual dubious articles, stressing questionable features: tortured writing style, citation of non-existent literature, and unacknowledged image reuse. Surprisingly, some websites offer to rewrite texts for free, generating gobbledegook full of tortured phrases. We believe some authors used rewritten texts to pad their manuscripts. We wish to raise the awareness on publications containing such questionable AI-generated or rewritten texts that passed (poor) peer review. Deception with synthetic texts threatens the integrity of the scientific literature.

Summary

The authors coin the term 'tortured phrases' for scientific terms garbled by word-for-word synonym substitution (e.g. 'counterfeit consciousness' for 'artificial intelligence') and trace them to one Elsevier journal, Microprocessors and Microsystems, where they concentrate. They show that journal's 2021 papers were peer-reviewed far faster than before, disproportionately came from mainland China, and had abstracts flagged as machine-generated by a GPT-2 detector at far higher rates than five comparison sets, and they document individual articles with reused images, fabricated citations, and unintelligible passages. They stop short of proving deliberate fraud and instead call on editors and publishers to investigate.

machine-generated · claude-sonnet-5

Outline

machine-generated · claude-sonnet-5

Claims

machine-generated · claude-sonnet-5

Key figure

Figure 5 — Side-by-side comparison showing that a suspect article's figure and its surrounding paragraph were copied, with the wording heavily and confusingly altered, from an unrelated water-leak-detector circuit tutorial on a hobbyist electronics website.

machine-generated · claude-sonnet-5

Glossary

Tortured phrase
A standard technical term mangled by substituting each word with a synonym, producing an odd phrase in its place, e.g. 'counterfeit consciousness' instead of 'artificial intelligence.'
Paper mill
An organization that manufactures and sells fake or low-quality manuscripts, authorships, or citations to researchers under pressure to publish.
GPT-2 Output Detector
A RoBERTa-based classifier, released alongside GPT-2, that estimates the probability a given piece of text was machine-generated.
Editorial assessment duration
The time elapsed between a manuscript's submission and its acceptance, covering screening, review rounds, and revisions.
Dvoretzky-Kiefer-Wolfowitz (DKW) inequality
A statistical bound used to build a confidence band around a distribution estimated from sample data, letting the authors claim two groups' score distributions differ with stated confidence.
Control set
A comparison group of texts, unrelated to the journal under suspicion, used as a baseline to judge whether the suspicious group's scores are unusually high.
PubPeer
A website where readers publicly comment on published papers to flag errors, misconduct, or other concerns after publication.
Retraction
A publisher's formal withdrawal of a paper's validity due to serious errors or misconduct, while the paper typically remains online marked as retracted.

machine-generated · claude-sonnet-5

Audience

Journal editors, publishers, and peer reviewers responsible for catching fraudulent or synthetic submissions; researchers in research integrity, scientometrics, and misconduct detection; science journalists covering publishing scandals.

prerequisites: Basic familiarity with how peer review and journal publishing work, Some awareness of language models like GPT-2/GPT-3 and what 'synthetic text detection' means, Comfort reading simple summary statistics and confidence intervals, though not required to follow the main argument

machine-generated · claude-sonnet-5

Open questions

machine-generated · claude-sonnet-5

Supplementary links

machine-generated · claude-sonnet-5

Cited by

Discussions

Related