Tortured Phrases: How Paraphrasing Tools Mangle Your Writing

Tortured phrases are the absurd word substitutions that automatic paraphrasing tools leave behind when they swap words one at a time without understanding meaning: “counterfeit consciousness” instead of “artificial intelligence”, “bosom peril” instead of “breast cancer”, “profound neural organization” instead of “deep neural network”. The term comes from research-integrity work by Guillaume Cabanac, Cyril Labbé and Alexander Magazinov, who found these fingerprints in thousands of published papers, many of which were later investigated or retracted. If you run text through a synonym-spinning paraphraser to lower a similarity score, tortured phrases are the residue it leaves, and automated screeners now hunt for that residue at scale.

This matters far beyond the papers that already got caught. Paraphrasing tools are marketed hardest at the people least equipped to spot their failures: writers working in English as a second language. This article explains where the term comes from, walks through the famous examples, shows why the tools produce them mechanically, and lays out what to do instead.

Where the term “tortured phrases” comes from

In July 2021, computer scientists Guillaume Cabanac and Cyril Labbé, together with Alexander Magazinov, posted a preprint describing “tortured phrases”: strange, unidiomatic substitutes for established technical terms that kept appearing in published research papers. Their hypothesis was simple and damning. Nobody writes “counterfeit consciousness” by accident. These phrases appear when someone feeds existing text through automatic paraphrasing or translation software to disguise its origin, usually to slip copied material past a plagiarism checker.

The name stuck because it captures what actually happens to the language. The text is not rewritten; it is tortured. Each individual word gets replaced by something a thesaurus considers close enough, and the meaning of the whole phrase dies in the process. In fields where terminology is fixed, this is not a style problem. “Breast cancer” is a defined medical term with decades of literature attached to it. “Bosom peril” is noise wearing a lab coat.

The scale turned out to be the real story. What began as a handful of oddities grew into a screening effort that has flagged thousands of published papers, triggered publisher investigations, and produced waves of retractions. One journal, Elsevier’s Microprocessors and Microsystems, became the emblematic case after investigators found hundreds of suspect papers concentrated in its special issues.

Tortured phrases: the famous examples

These substitutions were all found in real published venues by the tortured-phrases research effort. Read the left column, then the right, and you can reverse-engineer exactly what the software did.

Notice the pattern: every single-word swap is defensible in isolation. “Counterfeit” is a synonym for “artificial”. “Consciousness” overlaps with “intelligence” in some dictionary entries. The failure is at the phrase level. “Artificial intelligence” is not two words that happen to sit together; it is one unit of meaning, and a tool that only sees words cannot know that.

Why paraphrasing tools produce them

Classic paraphrasing tools, the kind usually called spinners, work exactly the way the examples suggest: look up each word, pick a replacement, move on. The tool has no model of what the sentence means, so it has no way to know that “random forest” names a machine-learning algorithm rather than unplanned vegetation.

The deeper problem is the objective these tools optimise for. A spinner’s whole job is to make text look maximally different from the source, because different-looking text scores lower on similarity checkers. But technical writing has words that must stay identical: names of methods, diseases, theorems, chemicals, legal concepts. A tool rewarded for changing everything will change precisely the words that carry the meaning. The objective is not slightly wrong for academic writing; it is pointed in the opposite direction.

Modern AI-based paraphrasers are better at grammar, which makes them more dangerous, not less. The output reads smoothly, so nothing jolts you into checking it. Terminology drift hides inside fluent sentences. And people who chain tools together, translating text into another language and back, or spinning an already-spun paragraph, compound the damage with each pass.

Why ESL writers are most at risk

I write in English as a second language, so let me describe the trap from the inside.

The pull towards these tools starts with fear. You get a similarity score you do not understand, or you have read about detectors falsely flagging human writing as AI, and the anxiety of being flagged, what we call flagxiety, makes a “rewrite it so nothing matches” button look like safety. Add the quiet belief many of us carry, that our English is not good enough on its own, and the spinner starts to look like a service rather than a hazard.

Then comes the part that is specific to second-language writers: we often cannot hear the absurdity. A native English reader laughs at “bosom peril” instantly, the way you would laugh at a mangled idiom in your own language. A second-language reader does something more careful and more dangerous: checks each word in a dictionary, finds each word individually fine, and accepts the phrase. The very diligence that makes ESL writers thorough is what lets tortured phrases through. Your ear for your native language would catch this in a second. Your ear for English may not, and the tool exploits that gap.

The result is a cruel irony. A student starts with honest writing, gets scared of a false flag, runs the text through a spinner as protection, and ends up with something worse than a false positive: genuine, verifiable fingerprints of tool misuse in work that began as their own.

How tortured phrases are detected

Cabanac, Labbé and Magazinov built the Problematic Paper Screener, a public tool that continuously trawls the published literature against a growing dictionary of known tortured phrases, with new fingerprints contributed by a community of volunteer sleuths. Flagged papers go to human review, and reviews have led to expressions of concern and retractions across multiple publishers.

It is worth understanding why this kind of detection is so much stronger than AI-writing detection. An AI detector gives you a probability, a statistical guess that reasonable people can dispute. A tortured phrase is checkable by anyone in thirty seconds: search the literature for “flag to commotion” and you will find no field that uses it; search for “signal to noise” and you will find half a century of engineering. There is no threshold to argue about and no model to second-guess. That is why tortured phrases end careers in a way detector scores rarely do; the evidence sits in plain text, in the author’s own submission.

And formal screeners are only half of it. Editors, reviewers and examiners now know the term. A single tortured phrase in a manuscript or thesis tells a reader everything they need to suspect about how the text was produced, even if no formal process ever starts.

The alternative: paraphrase meaning, not words

Real paraphrasing has never meant swapping words. It means understanding an idea, restating it in your own sentence structure, and citing the source. Under that definition, the terminology stays. Using the phrase “deep neural network” in a paper about deep neural networks is not plagiarism; it is competence. The parts of a sentence you own are the structure, the emphasis and the connective tissue, and those are the parts an honest rewrite changes.

A useful rule: nouns that name things, such as methods, diseases, algorithms and legal doctrines, are fixed. Everything around them is yours to rewrite. Any tool or habit that violates this rule is manufacturing evidence against you.

This is the principle Diglot’s paraphrasing is built on. It rewrites at the level of meaning, keeps technical terminology intact rather than “improving” it, and shows you every proposed change as a reviewable diff, so nothing enters your text without you seeing it. And because word-by-word conversion artifacts are exactly the failure mode ESL writers struggle to hear, we built a free Sounds-Translated Checker that flags phrasing which reads like mechanical conversion rather than natural English, the same family of artifact that tortured phrases belong to.

The deeper fix, though, is to stop playing the disguise game entirely. Spinning text is an attempt to hide a process. The stronger position is to prove one. Diglot’s editor can record your writing process as a signed, append-only chain of events and produce an Authorship Certificate: verifiable documentation that you wrote what you wrote, built while you work rather than reconstructed after an accusation. A writer with process evidence has no reason to touch a spinner, because the fear the spinner promises to solve is already answered.

A quick self-check before you submit

If any tool has touched your text, or you are simply unsure, run through this:

1. List your field’s core terms and search your document for them. Each one should appear in its standard form, every time.
2. Quote-search suspicious phrases. Paste an odd-sounding phrase into a search engine in quotation marks. An established term returns a wall of results; a tortured one returns almost nothing.
3. Never accept a synonym inside a multiword technical term. Single words can flex; fixed terms cannot.
4. Re-read every noun phrase in any passage a tool rewrote. That is where the damage concentrates.
5. Check that the rewrite actually reads better. Spun text is usually harder to read than the original, not easier; a readability check makes the comparison concrete.

Tortured phrases are what it looks like when a tool optimises for looking different instead of meaning the same. Your writing in a second language does not need to look different. It needs to mean what you intend, keep the vocabulary of your field intact, and carry proof that it is yours. That is a much easier standard to meet, and no spinner on earth can meet it for you.