Kidney Disappointment: The Two-Word Tell That Exposed an Industry

Imagine reading a medical paper and hitting the phrase kidney disappointment. The actual term is kidney failure. This is not a typo, and it is not a non-native speaker fumbling the language. It is the residue of software that swapped words to slip past plagiarism detection. Hundreds of papers carrying phrases like this are sitting in the published literature right now.

The dumbest possible attack, and it worked

The method is almost insultingly simple. Copy a paper wholesale. Run it through an automated paraphrasing tool. The tool has no idea what the sentences mean — it just walks a thesaurus and substitutes.

Failure can mean a collapse, but it can also mean letting someone down. So the tool reaches for disappointment. Dictionary-wise, defensible. The problem is that kidney failure is a single fixed medical term. Split it into two words and the meaning evaporates.

Academics call these tortured phrases. Guillaume Cabanac and colleagues at the Université Grenoble Alpes coined the label in a 2021 paper. The origin story is better than the term: they noticed computer science papers that kept saying counterfeit consciousness where any normal author would have written artificial intelligence.

The greatest hits are genuinely funny

Once you start collecting them, the list is hard to stop reading. Big data becomes colossal information. Cloud computing becomes haze figuring. Random forest turns into irregular timberland. Signal to noise ratio mutates into flag to commotion proportion. Mean square error ends up as mean square blunder.

Any first-year grad student would flag these in about one second. Which is the actual story here: these papers cleared peer review and were formally published. That means the reviewer didn’t read the manuscript, read it and shrugged, or never existed. All three answers point at the same diagnosis. The system has a hole in it, and the hole is load-bearing.

Paper mills are a real industry with real pricing

Behind this sits the paper mill — an organization that manufactures papers and sells them. Some sell authorship slots à la carte, priced by position. First author costs more. Third author is the budget tier.

The demand side is not mysterious. Plenty of countries and institutions still gate promotion, graduation, and grant renewal on publication count. Quality is hard to measure. Counting is easy. So an industry grew up to supply the count. It is a clean textbook case of what happens when you attach real money to a bad metric.

Cabanac’s group built a countermeasure: the Problematic Paper Screener, a public tool that scans published literature against a growing list of known tortured phrases. It has surfaced thousands of hits. Many of those papers have been retracted, and at least one journal was shut down entirely.

Then the language models arrived

Everything above is the story from 2021 through 2023. The uncomfortable part is what came next.

Tortured phrases were, from a detection standpoint, a gift. Crude synonym-swapping left a signature so loud you could grep for it. It was a fingerprint.

Modern LLMs leave no such fingerprint. They rewrite fluently. They know to leave kidney failure alone, because they distinguish fixed terminology from ordinary vocabulary. The detectable crudeness is gone.

There are worse signals. Multiple published papers have turned up with chatbot boilerplate sitting in the body text — lines like “As an AI language model, I cannot…” One had “Regenerate response” left in the introduction, a UI button label preserved for posterity in the scholarly record. Those are the ones sloppy enough to get caught. Nobody has any idea how many were not.

The part that should actually worry you

Contaminated papers get cited. Cited papers become the evidentiary basis for other papers. Those get cited too. A claim built on fabricated data burrows into the citation graph — the trust network that scholarship runs on — and takes root.

Now add the loop. Frontier models train on scholarly literature. A model trained on polluted papers goes on to help write more papers. The output feeds the input.

In medicine this stops being an abstraction. Meta-analyses pool individual studies into conclusions. Conclusions become clinical guidelines. Guidelines determine what happens to actual patients. Somewhere in that chain there may be a paper that wrote kidney disappointment.

The takeaway

We have arrived at the point where AI-generated problems require AI to detect. Human reviewers cannot read fast enough — the volume passed them years ago.

But detection is the wrong place to look for a fix. As long as people are evaluated by publication count, an industry will exist to supply publication count. Better detectors just escalate the arms race.

Kidney disappointment is funny. It was only ever funny because it was visible. The real question is how much of what you are reading right now isn’t.