\"Tortured Phrases\": Why Bad Paraphrasers Get Papers Retracted

In 2021, a Toulouse-based research integrity team published a list of phrases that had been appearing in computer science papers and that didn't quite make sense. "Counterfeit consciousness." "Haze figuring." "Profound learning." "Irregular esteem." "Bosom peril." Each was a thesaurus-substituted version of a real technical term — artificial intelligence, cloud computing, deep learning, random value, breast cancer. The papers using them had passed plagiarism checks but failed any human reader paying attention.

The team coined a name for these substitutions: tortured phrases. The original paper traced them to paraphrasing tools — both manual ones used by paper mills evading plagiarism detection and automated ones that some researchers were using on their own work. The discovery triggered a wave of retractions that continues into 2026. As of the latest count, more than 1,400 published papers have been retracted or flagged for tortured-phrase contamination, with most concentrated in engineering, computer science, and biomedical fields.

This guide explains what tortured phrases are, how they get produced, why they end careers and retract papers, the tools journals now use to detect them, and how to make sure your own paraphrasing workflow doesn't produce them.

What tortured phrases are

A tortured phrase is a recognized technical term that's been replaced with a synonym-substituted equivalent — usually one that's grammatically intact but semantically nonsensical to anyone in the field. The transformation typically happens word by word, without regard for whether the resulting phrase is something a domain expert would ever say.

A classic example from the original 2021 study: in a paper on "irregular timberland" (random forest), the authors discussed the model's "high precision" but didn't notice they had also written about "boundary trees" (decision trees) and "managed picking up" (supervised learning). The technical content of the paper was, in some sense, real research — but the prose had been processed by a tool that didn't know what any of these terms meant.

A typology of how they appear:

Single-term substitutions. Common technical terms replaced with their synonyms. "Artificial intelligence" becomes "counterfeit consciousness" or "fake brainpower." "Cloud computing" becomes "haze figuring" or "vapor processing." "Big data" becomes "huge information" or "enormous data."

Idiomatic phrase substitutions. Set phrases in the field replaced. "State of the art" becomes "condition of the workmanship." "Best in class" becomes "ideal in class." "Real-time" becomes "ongoing" or "constant time."

Acronym mishandling. Acronyms get replaced as if they were words. "MRI" becomes "attractive resonance imaging." "RNA" becomes "ribonucleic corrosive." "JavaScript" gets fragmented into "Java content."

Domain-specific substitutions. Field-specific terms get replaced with synonyms from a different domain. "Breast cancer" becomes "bosom peril." "Solar cell" becomes "sun-based cell." "Heart attack" becomes "coronary failure" (which is actually correct, but in a context where "heart attack" is the field-standard term).

The grammatical structure of the sentence usually survives. The semantic content is destroyed. A reviewer skimming might miss them; a reviewer with domain expertise notices immediately.

How they happen

Tortured phrases come from a few different sources.

Paper mills using synonym-substitution tools. This is the original concern that prompted the 2021 research. Paper mills produce fraudulent papers at scale, often by paraphrasing existing real papers and selling the resulting "new" papers to researchers buying authorship. To evade plagiarism detection, they run the source through aggressive synonym substitution. The output passes plagiarism checks (no exact strings match) and fails any domain expert's scan (the technical terms are wrong).

Researchers using free online paraphrasers without checking. This is the more common modern source. A non-native English speaker, a student under deadline pressure, or anyone trying to cut down word count runs their text through a free paraphrasing tool. The tool substitutes synonyms across the document. The author doesn't read carefully because the prose still reads grammatically — and submits a paper with field-specific terminology now wrong throughout.

LLMs in unusual prompt configurations. Modern language models like ChatGPT and Claude generally do not produce tortured phrases when asked to paraphrase, because they understand context. But certain prompting patterns can still trigger word-level substitution behaviors. Asking a model to "rewrite this with more variety" or "use synonyms throughout" sometimes produces tortured-phrase output, especially on technical content the model knows less well.

Translation pipelines through intermediate languages. Translating a paper from English to Russian to Chinese back to English (or any similar chain) can produce tortured-phrase patterns because each translation step substitutes word-level equivalents that don't recompose into the original technical terminology.

In our experience helping researchers, the second source — innocent researchers using bad paraphrasers — is the most common cause of tortured phrases in non-fraudulent papers. The author isn't trying to cheat; they're trying to improve their English or shorten a section. The tool destroys their terminology without telling them.

Why they get papers retracted

Tortured phrases are now treated as evidence of either paper-mill involvement or undisclosed AI use, both of which trigger retraction processes at most journals.

The reasoning has hardened over the past two years. Five or six years ago, a paper with strange phrasing might have been corrected — the editor would request the author fix the terminology and republish. Today, the same phrasing is treated as a sentinel for systematic problems. Even if the underlying research is sound, the presence of tortured phrases suggests the author either bought the paper, used an aggressive paraphraser to evade plagiarism detection, or didn't proofread their own work to a publishable standard. None of these are good explanations.

Specific journals have made the policy explicit. IEEE, ACM, Elsevier, and Springer have all updated editorial guidance in 2024-2025 to treat tortured phrases as grounds for retraction without requiring proof of intent. The 2026 Wiley editorial policy update added that "tortured phrases discovered after publication will result in retraction unless the author can demonstrate the substitutions resulted from a documented translation or editing process and not from paper-mill involvement."

The retraction also damages the author's reputation in ways that compound. Retracted papers stay on the author's record. Funding agencies check. Search committees check. Co-authors who weren't responsible for the phrasing get tagged in the retraction. Tortured phrases are one of the few editorial issues that can functionally end an academic career, especially for early-career researchers.

How journals are catching them now

The detection ecosystem has matured.

The Problematic Paper Screener. Developed by the same team that coined the term, the screener is a free online tool that searches PubMed and other databases for known tortured phrases. It's used by editors, peer reviewers, and integrity teams at journals to scan submissions and published papers. The tool maintains a regularly updated list of phrases — currently over 5,000 — and flags any paper containing them.

Editorial scanning before peer review. Several major publishers (IEEE, Elsevier, Springer) have integrated tortured-phrase scanning into their submission pipeline. Submissions are scanned at intake. Papers with hits are typically returned to the author with a request for explanation before peer review begins.

Post-publication monitoring. Tools like the Problematic Paper Screener also scan already-published papers. Hits trigger investigation by the publisher's research integrity team. The investigation may lead to expression of concern, correction, or retraction depending on what's found.

Reviewer awareness. Peer reviewers are increasingly trained to spot tortured phrases in their own field. Reviewer guidance from major journals now explicitly includes "scan for unusual technical terminology that may indicate paraphrasing-tool damage."

If your paper is going to be retracted for tortured phrases, you'll usually find out within 6-18 months of publication — often after the paper has been cited by others, which compounds the damage when those citations need to be tracked and notified.

How to avoid producing them yourself

A few habits prevent tortured phrases in your work.

Don't paste your entire methods or results section into a generic free paraphraser. This is the highest-risk action. Free paraphrasers optimized for "uniqueness" (i.e., evading plagiarism detection) typically use aggressive synonym substitution. They have no concept of which technical terms are field-standard and which are interchangeable. The output will have tortured phrases in the technical content.

If you must paraphrase technical text, use a citation-aware academic paraphraser. Tools like our paraphrasing tool are trained to preserve discipline-specific terminology and citation formatting. Field-standard terms — "deep learning," "random forest," "breast cancer," "structural equation modeling" — are preserved during the rewrite. Only the surrounding prose changes.

Read every paraphrased section before submitting. Specifically, scan for any technical term that looks unfamiliar or that you didn't write. If you see "counterfeit consciousness" where you would normally write "artificial intelligence," that's a tortured phrase the tool introduced. Restore the original term. Pairing the read-through with a pass through a tracked-changes AI proofreader makes substitutions easier to spot, since each edit appears as a discrete change you can reject.

Check translated text against the source. If you've translated a paper through any AI pipeline, scan the English version for any technical term that doesn't match what you'd write in your native field-appropriate English. Translation pipelines are a common source of tortured phrases, especially for English-to-Russian-to-English chains used to evade plagiarism.

Use the Problematic Paper Screener before submission. It's free. It takes about 30 seconds. It flags any known tortured phrases in your manuscript. If you've used a paraphraser at all, this check is worth running as the final pre-submission step.

Don't trust "uniqueness scores" as a proxy for quality. Tools that promise high "originality" or "uniqueness" scores typically achieve those scores through synonym substitution that produces tortured phrases. Plagiarism detection and good writing are not the same thing. A well-cited literature review with field-standard terminology may have moderate similarity scores because the field-standard terms are shared across papers — and that's fine.

Semantic paraphrasing vs synonym substitution

The distinction that protects your work is the difference between two fundamentally different approaches to paraphrasing.

Synonym substitution is what most free paraphrasing tools do. The tool takes a sentence, looks up word-by-word synonyms, and substitutes them. "The neural network achieved high accuracy" might become "The neural framework accomplished high precision." Sometimes the output is acceptable; often it produces tortured phrases. The tool has no understanding of what makes a phrase technically correct in your field.

Semantic paraphrasing is what good academic paraphrasers do. The tool understands the meaning of the sentence and rewrites it while preserving the technical terms that are field-standard. "The neural network achieved high accuracy" might become "Our neural network reached high accuracy on the benchmark." The terminology is preserved because the tool recognizes "neural network" as a field-standard term, not a phrase to be synonymized.

The difference is structural, not cosmetic. A tool built on semantic paraphrasing will preserve "breast cancer" because it recognizes the medical-domain context. A synonym-substitution tool will replace it with "bosom peril" because, at the word level, "bosom" and "peril" are dictionary synonyms of "breast" and "cancer."

You can usually tell which approach a paraphrasing tool uses by testing it on a paragraph dense with technical terminology. Paste a methods section into the tool. Read the output. If the technical terms survived (deep learning, random forest, structural equation modeling, breast cancer), the tool is using semantic paraphrasing. If you see novel phrases (profound learning, irregular timberland, primary condition demonstrating, bosom peril), the tool is using synonym substitution and should not be used on academic content. For more on what to look for in an academic-grade paraphraser, see our comparison of paraphrasing tools that preserve citations.

See the Full Paraphrasing Tool

Citation-aware academic paraphrasing that preserves technical terminology. Free tier includes every feature.

Frequently asked questions

Q: How can I tell if my paper has tortured phrases before submitting?

Run it through the Problematic Paper Screener (free at the Cabanac et al. project site). The tool checks your manuscript against a list of over 5,000 known tortured phrases. Hits are flagged with the corrected term. You can also manually search your paper for any technical term that doesn't match your normal vocabulary in the field. If you've used a free paraphrasing tool on the manuscript, the highest-risk sections are usually methods and results, where technical terminology is densest.

Q: What should I do if I discover my submitted paper has tortured phrases?

Contact the editor immediately, before peer review concludes. Explain what happened (you used a paraphrasing tool that introduced the substitutions). Provide a corrected manuscript. Editors generally treat proactive disclosure very differently from later discovery. Most will accept a corrected resubmission if you act before the issue is detected independently. If the paper has already been published, contact the journal's research integrity office to request a correction. The earlier you act, the lower the chance of formal retraction.

Q: Are LLMs like ChatGPT and Claude safe to use for paraphrasing?

Modern LLMs are generally much better than dedicated free paraphrasers at preserving technical terminology, because they understand context. They are not immune to tortured phrases, however. Specific prompting patterns ("use synonyms," "rewrite for variety," "make it more unique") can trigger word-level substitution behaviors. If you use an LLM for paraphrasing, prompt it to "preserve all technical terminology and citation formatting" explicitly, and always verify the output against your source. For high-stakes paraphrasing on a publishable manuscript, a citation-aware academic paraphraser is safer than a generic LLM prompt.

Q: How does the Problematic Paper Screener stay up to date?

The team behind it (led by Guillaume Cabanac at the University of Toulouse) maintains an open list of phrases collected from forensic analysis of retracted and flagged papers. The community contributes new findings. As paper mills evolve their substitution patterns, new phrases get added. The list currently exceeds 5,000 entries and grows monthly. If you discover a tortured phrase in your own field that isn't on the list, you can contribute it — the screener becomes more useful as the community adds field-specific terminology.