Tortured phrases

Guillaume Cabanac, Cyril Labbé, and Alexander Magazinov coined the term in 2021 after finding published papers riddled with bizarre rewordings of standard terminology: "counterfeit consciousness" for artificial intelligence, "profound learning" for deep learning, "colossal information" for big data. Because no researcher writes those phrases on purpose, they became a searchable fingerprint for screening the literature for machine-paraphrased text.

How does "counterfeit consciousness" happen? Thesaurus-style paraphrasing tools — spinners — replace words one at a time with near-synonyms, without knowing that some word pairs are fixed terms of art. Artificial → counterfeit and intelligence → consciousness are each defensible swaps in isolation; applied to a two-word technical term they produce nonsense. The tools are typically used to push text past similarity score checks: the strings no longer match, so the software goes quiet — while the writing gets objectively worse.

For ESL writers the lesson is direct, because paraphrasing tools are marketed at them hardest. Word-swap paraphrasing is the failure mode, not the feature: it preserves the source's skeleton (the thing that makes patchwriting detectable) while corrupting the vocabulary (the thing that made the sentence correct). Real paraphrase restates understanding in a new structure. A practical test: if a tool's output contains a phrase no one in your field would ever write, that phrase is the tell.