The Slang That Comes from the Machine

A few days ago I wrote about how much things „carry“ these days. Packages carry costs. Tools carry functions. Analyses carry metrics. A week later it was „maturity level“: everywhere on LinkedIn, someone is bringing something to maturity, as if a company were an apple. Both felt like a linguistic irritation, though that alone proves nothing. Now three papers have landed on the table that have investigated exactly this irritation. One study counts words across 26 million PubMed entries. Another group of researchers counts them across hundreds of thousands of hours of podcasts and YouTube videos. The third paper explains where the jargon comes from and why it is embedding itself ever more deeply in our language. Together they do something a newsletter cannot do on its own: they show that „carries“ and „maturity level“ are not quirks. They are the German equivalents of the English words „delve,“ „underscore,“ and „meticulous.“

What the numbers show, before intuition kicks in

Kentaro Matsui searched PubMed from 2000 to April 2024, looking at 118 words and phrases that users on Reddit, X, and various forums had flagged as typical of ChatGPT. As a control, he included 75 entirely ordinary academic phrases. The result: 75 of the suspect expressions spike sharply in 2024. At the top are „delve,“ „underscore,“ „meticulous,“ and „commendable.“ The rise does not begin on November 30, 2022. It is already visible around 2020 and continues to grow. The control phrases remain comparatively flat.

This is how „machine language“ develops. Linguistically speaking, rare words or loan words are used more and more frequently, eventually becoming part of the training material for large language models. The more often these words appear in the memory of language models, the more frequently they turn up in texts written by the machine.

Hiromu Yakura of the Max Planck Institute for Human Development noticed the same suspicion in himself. He suddenly found himself saying „delve,“ even though the word had never been characteristic of him. His team analyzed more than 700,000 hours of audio: hundreds of thousands of podcast episodes, hundreds of thousands of YouTube videos, before and after ChatGPT. The same model words appear measurably more often in spoken language afterward. Not only in read-aloud scripts. Also where people speak freely.

In German, coverage of this research lists words including betonen, tiefgreifend, bahnbrechend, revolutionieren, and spiegeln as typical machine language. That is the bridge between what Matsui counts in abstracts and what Yakura measures in voices — and what arrives in our own writing as „carries“ and „maturity level.“

Where the new language comes from

Alexandra Frye laid this out in a piece for the APA in exactly the terms I would have wanted when I first stared at „carries“ in a SaaS newsletter. The early models were trained on mountains of text from content farms and freelance platforms — cheaply produced, search-engine-friendly prose. Strong English, rigid style guides, smooth transitions, little originality. The models did not only learn facts. They learned a tone: professional, balanced, mildly ceremonious, and short on originality. That is why the default language of AI-generated text sounds so clean and so interchangeable.

„Delve“ is not an accident. It is a word that simulates authority without becoming specific. „Meticulous“ and „commendable“ work the same way. Frye calls this AI slang — referred to in English as an AI accent. It is a watermark for those who listen closely, and it is likely part of how tools like Claude are detected.

„Carries“ works in exactly the same way. Five English verbs — cover, support, bear, sustain, contribute — mean five different things in German. The model throws them into one pot because „trägt“ appeared most often alongside all of them in the training data. Language as a majority vote. „Maturity“ becomes „Reifegrad“ because the assembly instructions are in English and the parts are in German. It is not wrong. It is clunky. And because it appears everywhere, you eventually notice that a distinct language of its own is at work.

All three texts describe the same entanglement, only with different vocabulary. Models favor certain words. People read and hear them. People take the source to be authoritative — once it was the professor, the editor-in-chief, the specialist colleague; today it is often the tool. People adopt the words, often without noticing. The words turn up in new texts, in talks, in podcasts. The next generation of models is trained on them again.

This has always happened. Goethe and Schiller shaped the language and expression of their contemporaries just as OpenAI, Grok, and Claude are doing today. Yakura speaks of a cultural feedback loop. Frye speaks of a linguistic feedback loop and of the danger that the shared pool of linguistic possibilities is narrowing. Matsui is more cautious: ChatGPT has accelerated a trend that was already there. A parallel universe without ChatGPT does not exist, he writes. Even so, what we can already see is enough.

Anyone who writes attentively notices the clunky spots in AI-assisted texts. Anyone who was always inclined to mix German and English will fall into the rabbit hole of loan words and unknowingly mark themselves as an AI user. The loan-word detector that Klaus and I built is not an AI police force. It is a brake. It asks: what was meant in English, and what would I say if I had to say it myself?

The Thermomix stays a Thermomix

The distinction I drew previously still holds. The father of a family on a Tuesday evening needs the Thermomix to put a decent Bolognese on the table. He is not aiming for a Michelin star. Senior professionals on LinkedIn who replace their communications department with a language model are aiming for a star but serving linguistic Thermomix. You do not taste it in the dash. You taste it in the word that nobody used before.

Matsui writes that ChatGPT texts in medicine are fluent and logical, but more general, less specific. That is precisely what „carries“ is: a commitment without edges. To cover, to take on, to support, to contribute — those are four different responsibilities. Texts live by these distinctions. Anyone who flattens them because the language model has already done a read-through is replacing the editor with statistics.

Frye invokes the principle of least effort: we reach for the language that is closest to hand, especially when the tool seems authoritative. Creativity requires deviation. Models smooth deviation out. Yakura warns that the frequency of certain words can change how we describe situations and construct arguments. This is not stylistic nitpicking. It is the claim that word choice shapes how we access the world.

What remains when you listen closely

The three papers do not support my two earlier pieces by counting „carries“ or „maturity level.“ They do not do that. They support them by making visible the mechanism that produces these words.

First: models have favorite words. In English, these have already been measured. In German, you recognize them through calques and all-purpose verbs. Where others already have data, we currently have only a feeling — and there we find ourselves back with Goethe and Schiller.

Second: the rise begins two years before the launch of AI, already around 2020, and afterward becomes as steep as a ski jump. Anyone writing with the machine in 2026 has long since been breathing the same air, even without ever having opened the tool. The editor has been breathing it. The colleague who drafts the newsletter has too. You do not need to operate the machine to be standing in its exhaust.

Third: spoken language follows. What appears in abstracts soon turns up in podcasts. What appears in newsletters slips into the sentence you call out to your neighbor on the staircase.

Fourth: recognizability fades the more people adopt the machine’s slang. Evans, quoted in coverage of the Yakura study, says in effect: at this stage, word distribution is the right indicator. Later it will be harder.

That is why I stay with the simple discipline that Frye formulates at the end — and that I would put in German like this: notice. Notice what you take in. Notice what you repeat. Notice whether the sentence is still your own or whether the machine’s gravitational pull has taken over.

Since reading these texts, I have not used „delve.“ I keep striking out „carries.“ I only use „maturity level“ when an actual apple is involved.

The family man’s meal is allowed to taste of Thermomix. The executive who is bringing his organization to maturity should first check what he actually wanted to say.

This English version was created with the help of AI (translated from the German original) and may differ slightly from a professional human translation.