Did automated text tools produce suspicious phrases in one journal?

The paper's central claim is that a specific writing defect marks a cluster of questionable science. "Tortured phrases" are "unexpected weird phrases in lieu of established ones," such as "counterfeit consciousness" for "artificial intelligence" or "flag to clamor" for "signal to noise." The authors find them concentrated mainly in Computer Science, most heavily in Microprocessors and Microsystems, an Elsevier journal. Across its 1,078 full-length articles in volumes 56–83 (February 2018 to June 2021), they report "a sudden shortening of editorial assessment for volumes published in 2021," with most articles published after an assessment "surprisingly short," China and India overrepresented, and blocks of articles sharing identical submission and acceptance dates. They estimate around 500 questionable articles accepted, counting 389 in volumes 80–83 and 225 "in press" as of June 25, 2021.

The authors' mechanism is a hypothesis, not a finding. They argue the phrases were "coined by misused natural language processing (NLP) tools: automatic translation, automatic re-writing or even automatic generation of text," because a skilled scientist would not invent nonstandard terms for well-known concepts, nor switch to them while citing the standard literature. Their detector screen supports the hypothesis only partly. They ran the GPT-2 Output Detector on abstracts and found a concentration flagged as synthetic, then added that "texts flagged as synthetic by the GPT detector might be scientifically sound," and read individual papers to check. The case reviews record reuse of text and images without acknowledgement, references to non-existing literature, and sentences "for which we failed to infer any meaning." As of 17 June 2021, none of the 1,078 papers had a PubPeer comment.

Set against Does polished writing actually signal better quality work?, this case runs the other way. The marker here is clumsiness, not polish, and the papers passed editorial review anyway, which is the authors' point about deficient peer review. Set against Can people reliably spot content made by AI? and Can readers tell LLM abstracts from human ones?, the paper tests no human readers at all. Its evidence is a lexical fingerprint, an automatic detector and publication metadata. That makes it narrower than Do AI slop accusations actually detect AI text?, which concerns general prose features. A fixed list of odd phrases catches a failure those features miss. It sits closest to the style-quality axis in What dimensions make text feel like AI slop?, though the authors do not use that vocabulary.

What the excerpt does not establish is the causal chain. The authors describe their evidence as "indication – if not evidence" and say "without any definitive proof." They did not measure whether NLP tools produced the phrases, and they say the shortened timelines "may reflect poor or deficient editorial assessment" without ruling out other causes. The GPT-2 detector belongs to an older model generation than the advanced models they hypothesize. The prevalence figure of 4.29 papers per million is one they cite from Cabanac & Labbé and Van Noorden, not one they verified, and their image screening was "not performed systematically." The implication, at the strength the evidence allows, is that the tortured-phrase cluster and the editorial-timeline anomaly justify an audit of one journal and a wider search of the literature. They do not yet give a measured rate of AI-written science.