'Counterfeit Consciousness': What an AI Rewrite Does to the Words You Search By

Across the 8,282 entries cataloging an established scientific term destroyed by automated rewriting, 93.8 percent of those terms are multi-word phrases. Single words mostly survive. Phrases do not. That distribution, counted today from the researchers' own export, is the finding — and multi-word phrases are the handles your future search has.

The literature calls the wreckage a tortured phrase. The naming paper is Guillaume Cabanac, Cyril Labbé and Alexander Magazinov, Tortured phrases: A dubious writing style emerging in science, posted to arXiv on 12 July 2021. It is a preprint — version 1, no journal reference — and it should be read as one.

Its abstract states the concept plainly: "Our study introduces the concept of tortured phrases: unexpected weird phrases in lieu of established ones, such as 'counterfeit consciousness' instead of 'artificial intelligence.'"¹ The peer-reviewed measurement arrived later, at the ACL's WIESP 2023 workshop.² Between them sits a public catalog of every destroyed term the researchers have found, and that catalog is what this post downloaded and counted.

The study, briefly

The 2021 preprint named a pattern that had been sitting in plain sight in published journals: "Kidney disappointment" for kidney failure, "bosom peril" for breast cancer, "Fake neural organizations" for artificial neural networks.³ The documented cause is not carelessness. It is software.

Cabanac, Labbé and Magazinov describe the mechanism in section 2 of the preprint: "well-known and well-established scientific terms were replaced by unconventional phrases. In a typical case, a word-by-word synonymical substitution is applied to a multi-word term."⁴ The tool class is named in the same paper: "Surprisingly, some websites offer to rewrite texts for free, generating gobbledegook full of tortured phrases."⁵

Why anyone would run such a tool is the least mysterious part. Writing in The Conversation, Cyril Labbé, Frederik Joelving and Guillaume Cabanac put it directly: "They typically result from using paraphrasing tools to evade plagiarism-detection software when stealing someone else's text."³

The definition the researchers settled on is compact enough to carry: "A tortured phrase is an established scientific concept paraphrased into a nonsensical sequence of words."⁶

What they actually measured

The screening runs continuously. The Problematic Paper Screener, built by the same group, "combs through 130 million scientific publications every week" and has been "instrumental in more than 1,000 retractions."³ The catalog it screens against has grown from 473 phrases in January 2022 to 8,282 as of its 13 August 2026 update — a seventeenfold rise, both numbers first-party.⁷⁸

The method behind the first sweep was ordinary and reproducible. The team screened a sample of about 140,000 abstracts of papers published by one large publisher, then went looking for the phrases elsewhere.⁹ By January 2022 they had found tortured phrases in 3,191 peer-reviewed articles.⁷ By 13 October 2023, the peer-reviewed WIESP paper by Eléna Martel, Martin Lentschat and Cyril Labbé reported 11,945 flagged papers, 1,278 of them published in 2023 alone.¹⁰

The most direct evidence in the literature is one sentence: "A strong piece of evidence is that one can reproduce most tortured phrases by feeding established terms into paraphrasing software."¹¹ The mechanism is not inferred. It is re-run on demand.

Two honest notes on the method. The public detector table is deliberately conservative — it lists only "Articles matching 5+ tortured phrases or 1+ obvious tortured phrases (e.g., sign to clamor)," which the site describes as "a 3% subset of all articles found with at least one tortured phrase."¹²

Automated detection of new tortured phrases is also imperfect in the direction that matters. The WIESP team reports "an interesting recall of 0.873, showing a good capability to detect new tortured phrases, but a precision of 0.615."¹³ Roughly two in five machine-proposed phrases are wrong, so domain experts still filter the output.

Where the damage lands: 93.8 percent

Each row of the catalog pairs a tortured phrase with the established term it replaced, which makes it a record of what was destroyed. Download the 8,282-row export and count word lengths, and the distribution is lopsided: 7,771 rows — 93.8 percent — record a destroyed term two words or longer. Only 511, or 6.2 percent, are single words.¹⁴

# The Problematic Paper Screener's own fingerprint export, 2026-08-13 19:27
curl -sL "https://dbrech.irit.fr/pls/apex/f?p=9999:5::IR%5Btortured%5D_CSV::::" -o tort.csv
wc -l tort.csv            # => 8283  (header + 8,282 rows)

python3 - <<'PY'
import csv, re, collections
rows = list(csv.DictReader(open('tort.csv', encoding='utf-8')))
w = lambda s: len([x for x in re.split(r'\s+', s.strip()) if x])
c = collections.Counter(w(r['Expected Text']) for r in rows)
multi = sum(v for k, v in c.items() if k >= 2)
print(len(rows), multi, f"{multi/len(rows)*100:.1f}%")     # => 8282 7771 93.8%
PY

The full shape: 511 destroyed terms of one word, 4,145 of two, 1,736 of three, 1,265 of four, 381 of five, 244 of six or more.¹⁴

That single-word bucket is smaller than it looks. In 182 of those 511 rows the researchers logged a tortured phrase but never identified the original term, so the cell reads ??? rather than a word. Count only the 8,100 rows that name what was destroyed and the multi-word share rises to 95.9 percent.¹⁴ This post keeps the conservative 93.8. The swap is phrase-for-phrase rather than word-for-word at the other end too — the replacement string is itself multi-word in 8,109 of 8,282 cases, or 97.9 percent.¹⁴

This is the mechanism the WIESP authors describe: "By performing a 'word by synonyms' replacement, the first generation of spinners would destroy the meaning conveyed by these typical collocations."¹⁵ A collocation is a phrase whose meaning is not the sum of its words. "Signal to noise" is one. "Big data" is one.

Substitute inside a collocation and it stops pointing at the thing. As the researchers put it in the Bulletin: "Replacing a word by its synonym in lay language may lead to a different scientific meaning."¹⁶

One methodological caution, because the export invites the mistake: the catalog's retrieved-papers column is a retrieval count, not a count of confirmed tortured papers. A single paper matches many fingerprints, and 653 fingerprints — 7.9 percent — are compound queries requiring a second term to disambiguate.¹⁴ Summing that column produces a large, double-counted number that means nothing. The defensible counts are the ones above.

Why this matters for your own notes

Your notes carry established terminology too. The project code name. The client's exact phrasing. The error string you pasted at 2 a.m. Those are multi-word tokens, and they are the only handles a future search, backlink or grep has. A rewrite that swaps them leaves the note more readable and makes it unfindable.

Here is that claim as a receipt rather than an assertion. Two files, one rewrite, using substitution pairs documented in the catalog itself — flag to clamor for signal to noise, neural organization for neural network:¹⁴²

printf '# Meeting 2026-08-14\n\nWe agreed the retrieval layer must keep the signal to noise ratio above the\ncurrent baseline before we enable the neural network re-ranker.\n' > original.md
printf '# Meeting 2026-08-14\n\nWe agreed the retrieval layer must keep the flag to clamor ratio above the\ncurrent baseline before we enable the neural organization re-ranker.\n' > rewritten.md

wc -w original.md rewritten.md
#  26 original.md
#  26 rewritten.md

grep -c "signal to noise" original.md rewritten.md
#  original.md:1
#  rewritten.md:0

grep -c "neural network" original.md rewritten.md
#  original.md:1
#  rewritten.md:0

Twenty-six words in, twenty-six words out. Three changed tokens. Both search handles gone.¹⁷ Nothing about the note looks damaged — a reader skimming it would call it clean. The damage is only visible from the future, at the moment you type the phrase you are sure you wrote.

That receipt is a constructed example, not a study of anybody's vault. Be precise about which part is measured: the 93.8 percent is measured, the substitution pairs are documented, the grep behavior is arithmetic. What no published study we could obtain measures is how often this happens to personal notes. Naming that gap is cheaper than papering over it.

The consequence runs through the rest of the toolkit. If literal search is enough for a small vault — and it usually is — then the precondition that argument silently assumes is that the literal string is still there.¹⁸ Structure decides what a retriever finds inside a note; how you split a note works one level above this one.¹⁹

This is also a different failure from meaning drift. An AI summary over-generalizes what you meant; a rewrite changes the strings you would use to get back to it.²⁰ Meaning is one loss. The handle is another.

What it costs to undo

Restoration is a live research problem, and the numbers are not encouraging. A December 2025 preprint by Agniva Maiti, Prajwal Panth and Suresh Chandra Satapathy reports that "zero-shot baselines fail completely (0.00% restoration accuracy)," while their purpose-built retrieval system "achieves a restoration accuracy of 23.67%."²¹ Under a quarter — and only with a corpus to search.

The reason a language model alone scores zero is the interesting part, and the authors state it: tortured phrases "are often grammatically valid synonyms," so an isolated model has no basis to object. "It lacks the ground truth to know that 'big data' is a fixed terminological entity in Computer Science."²² The phrase reads fine. It is simply the wrong phrase.

Their better result depends entirely on having the source: "The restoration module is strictly limited by the contents of the reference corpus."²³ Translated to a desk, that is the whole practical rule. If you kept the original, you can recover the term. If the rewrite is the only copy, you are asking a model to guess which fixed expression a synonym was standing in for — and that is the task that scores zero.

Before you click "improve this"

Five moves, none of which require refusing the rewrite. The discipline is not abstinence; it is about what survives the edit. Each one exists because a specific recovery path fails without it. The first is the only one that is not optional, because every other move depends on the original still being on disk.

1. Copy the file first, under a name you will recognize in six months. Restoration without the source is the 0.00 percent case.²¹ With the source it is a diff.
2. Write down the terms that may not change, before the rewrite runs. The names, error strings and fixed expressions in that note are your retrieval keys; a glossary file your AI can read is the durable version of this move.²⁴
3. Diff the nouns, not the tone. Tone changes are the point of a rewrite. Noun-phrase changes are what your eye forgives.
4. Grep your handles afterwards. grep -c "the exact phrase" note.md before and after takes ten seconds and returns a number, which is more than a read-through returns.
5. Never let the rewritten version be the only version. We have said before that letting a model rewrite your raw sources is how errors become permanent.²⁵

None of this forbids the rewrite. It refuses to let the rewrite become the record.

The caveats

Three limits sit under everything above, and each one is load-bearing. The measured population is published scientific abstracts and papers, not personal notes. The documented cause is plagiarism-evading paraphrase software, not an assistant tidying your meeting notes. And the field's most-cited paper is a preprint.

On scope. Nothing in this literature shows that a language model does this to your notes. What it shows is large, named and repeatable: this is what automated rewriting has been measured to do to fixed terminology, across a catalog of 8,282 entries.⁸ Your notes have fixed terminology. The transfer is an argument, and it is ours.

On generation. The anchor evidence describes the 2021 spinner generation, and the WIESP authors concede the point themselves: "While it's possible that with the context of 2023, Large Language Models can perform paraphrasing of higher quality than spinners, it's crucial to note that these papers have already been published and remain accessible."²⁶ Modern paraphrase is more fluent. Fluency is not the same as leaving a fixed expression alone.

On peer review. Tortured phrases: A dubious writing style emerging in science is a preprint that was never formally published, and calling it peer-reviewed would be an easy error to make. The peer-reviewed anchor here is the WIESP 2023 paper in the ACL Anthology.² The December 2025 restoration work is a preprint too, and is labeled as one everywhere it appears above.²¹

Two more complications, both of which make the picture better rather than worse. Machine translation is a genuine competing explanation for some individual specimens in the wild, and readers of the discussion that surfaced this topic argued it at length. The researchers' documented cause is paraphrase software and the reproducibility test supports them, but "every example is a spinner" would be an overclaim.¹¹

And 23.67 percent restoration is not zero. The honest phrasing is mostly unrecoverable, and only with the original indexed, not impossible.²³

Frequently asked questions

Six questions people actually ask about automated rewriting and their own notes, answered from the sources above rather than from general productivity advice. Where the honest answer is a concession, it is a concession. Where the evidence stops short of the question, the answer says so instead of extending the finding past what the researchers measured.

Is it safe to let AI rewrite my notes?

It is safe if you keep the pre-rewrite original. The restoration research is unambiguous on the alternative: a zero-shot model recovers 0.00 percent of destroyed terms, and a purpose-built retrieval system recovers 23.67 percent, and only when the source text is indexed.²¹²³ With the original file on disk, a bad rewrite is a diff. Without it, it is a guess.

Why can't I find a note I know I wrote?

One documented cause is that the words changed. A rewrite that substitutes inside a multi-word term leaves the note the same length and the same apparent quality while removing the exact string you would search for. In a 26-word test note, three changed tokens took grep on both search handles to zero.¹⁷

Could this be a translation issue?

Sometimes, and that is worth conceding. Machine translation can produce similar-looking artifacts. But the researchers' documented cause is paraphrasing software, and their reproducibility test is direct: most tortured phrases can be reproduced by feeding established terms into a paraphraser.¹¹

Isn't it more likely to be changing words to disguise plagiarism?

That is exactly the stated cause, not an alternative to it: "They typically result from using paraphrasing tools to evade plagiarism-detection software when stealing someone else's text."³ The motive is plagiarism evasion. The mechanism — word-by-word substitution inside a fixed multi-word term — is the part that transfers to any other use of the same class of tool.⁴

What is a content spinner?

A spinner is a tool that rewrites text by substituting synonyms. The WIESP definition is precise: "We define a tortured phrase as an expression resulting from the use of a spinner on a well-established scientific expression with a specific and fixed meaning."² The 2021 preprint observes that "some websites offer to rewrite texts for free, generating gobbledegook full of tortured phrases."⁵

Do large language models produce tortured phrases?

The published catalog is built on pre-LLM spinner output, and no study cited here demonstrates that a modern model does the same thing at the same rate. What the 2025 restoration work shows is that the recovery problem persists in the LLM era: a language model has no ground truth telling it which multi-word expressions are fixed, which is why zero-shot restoration measures 0.00 percent.²¹²²

The words are the address

Readability is a property of the sentence. Findability is a property of the words. A rewrite trades the second for the first, quietly and without leaving a mark, and only one of the two is how you get the note back. In the catalog documenting this, 93.8 percent of destroyed terms were phrases.

And phrases are what you type when you go looking.

This post stands on work by Guillaume Cabanac, Cyril Labbé and Alexander Magazinov, who named the phenomenon and built the public catalog it was counted from, and on the WIESP measurement by Eléna Martel, Martin Lentschat and Cyril Labbé.

MNMNOTE keeps notes as plain Markdown files on your own device, which is what makes a pre-rewrite copy and an after-the-fact diff two ordinary file operations: mnmnote.com.