Marginalia

"Shape," "Table Stakes," and Other Borrowed Words: Is AI Changing How We Talk?

Language & LLMs

I owe my love of language to Tolkien. It carried me through adolescence, faded over the years as my gifts with technology and the career I built with them took over, and I assumed it was gone for good. The spread of language models has rekindled it, and these days I catch myself paying close attention to individual words and grammar again.

That’s how I started noticing colleagues at work using “shape” where it doesn’t quite belong: asking about the shape of a dataset, an API response, or a requirements doc, with the word stretched to cover the structure or schema of nearly anything. Commentators call a product feature “table stakes” on a loop. Essayists and the YouTube explainers I watch reach for the same small shelf of terms with a regularity that begins to feel like déjà vu. None of these words is strange on its own. What snags me is that they’re also among the most recognizable tells of LLM prose, and I now run into them in places where no model is visibly involved.

I’m not the only one. A recent meme in the office “AI bemusement” chat showed the two most popular providers’ mascots sitting at a desk, surrounded by the ephemera and iconography of LLM catchphrases like “say the word” and “belt-and-braces.”

The obvious suspicion is that I’m the instrument at fault. I read a great deal of machine-generated text, I’ve had its tics memorized for a while, and anyone in that position can find those tics everywhere. That’s the frequency illusion, or plain confirmation bias. I carry plenty of my own skepticism. Still, the intuition is testable, and several research groups have run versions of the test on large corpora. Their results are worth a look before I speculate.

After what felt like the fortieth time noticing this, two questions settled in my head. Is any of it real, or am I just primed? And if it is real, does it stop at borrowed words, or could enough LLM use nudge human speech toward the models’ training distribution, and would that take one generation or several? Underneath both is the version that worries me most, in which the loop goes recursive and future models train on language that earlier models already bent. So the question I keep circling back to is what drives language change, and whether any of it pushes back.

Is it real?

The hardest numbers I’ve found come out of medicine. Kobak and colleagues went through more than 15 million PubMed abstracts from 2010 to 2024, using a trick borrowed from epidemiology: project each word’s expected frequency from its pre-2023 trend, then measure the excess afterward, the way public-health researchers estimate excess mortality. In 2024, “delves” appeared roughly 28 times more often than its own history predicted, “underscores” about 14 times, “showcasing” about 11. From the aggregate excess, the authors infer that at least 13.5 percent of 2024 abstracts passed through an LLM at some point, with the estimate running from under 5 percent to over 40 percent depending on the field, country, and journal. I’d love to know what that number looks like now, two years deeper in.

One comparison in that paper stuck with me. The pandemic also produced a surge in biomedical vocabulary, but it was almost all content words, like “respiratory” and “remdesivir,” that tracked new things in the world. The post-2023 surplus is almost all style words. Nothing new entered the literature that would call for them; authors, or their tools, started picking different words for the same ideas. That’s a shift in register rather than reference, which is exactly where a model’s stylistic preferences would leave a fingerprint.

Abstracts, of course, are the texts most likely to have gone through a model in the first place. Speech is harder to explain away. Yakura and colleagues at the Max Planck Institute for Human Development analyzed 737,083 hours of conversation drawn from 824,634 podcast episodes, screened down to unscripted talk, and found an abrupt post-ChatGPT rise in the words models preferentially generate: “delve,” “showcase,” “boast,” “intricacies,” “meticulous.” Their reading is that systems trained on human language are now returning their own statistical signature to it.

Before I build anything on that, I should say what these studies don’t show. A well-timed jump in the data doesn’t tell you what caused it. Both papers find a structural break coinciding with ChatGPT’s release; neither one watches a person pick up a word from a model. Several routes fit the same aggregate data: direct accommodation to machine prose, indirect diffusion through writers who themselves use LLMs, and people reading aloud material a model helped draft. The routes imply different things about how durable the change will be. The corpus can’t tell them apart.

The measured words also aren’t the ones that started me on this. “Delve” has a documented excess. “Shape” and “table stakes” don’t, to my knowledge. Until somebody does the broader sampling and analysis, my observation stays an anecdote. Or until somebody asks Gemini to do it.

Both of my words have perfectly good histories that never come near a model. “Table stakes” is from poker; the phrase apparently migrated into management writing decades before any language model existed, and the explanatory-video genre I watch draws that idiom straight from business and consulting prose. If models are fond of it, that may reflect how common it already was in their training data, with humans and machines drinking from the same well and nobody influencing anybody. “Shape” has a similar pedigree. Software developers were describing the “shape” of data long before LLMs: React’s PropTypes library has offered a shape validator for years, and “the shape of the data” is ordinary shop talk among programmers. Models trained heavily on code and technical writing would inherit the usage, so its drift into general professional speech could come from the models, from the growing number of people who work near software, or from both. Disentangling convergence from transmission is hard, and I’d be skeptical of anyone making confident claims either way off a handful of examples.

So LLM-linked lexical change looks real for a measurable set of words. The writing evidence is solid and the speech evidence suggestive, but it remains stubbornly unproven for the two words I set out to explain.

Grandpa and Grandma had a way of speaking

Whether this generalizes beyond vocabulary is a question sociolinguistics has been asking for sixty years. Labov’s studies of New York City speech in the 1960s established the trick known as apparent time: compare speakers of different ages at a single moment, on the assumption that each cohort’s speech preserves something of the period in which it was acquired. The method leans on a distinction between generational change, where successive cohorts use an innovative form at progressively higher rates, and age-grading, where individuals shift their usage at particular life stages while the community’s overall pattern stays put.

Different aspects of language turn out to be plastic to different degrees. Phonology and core grammar mostly settle by late adolescence, though Sankoff and Blondeau, following Montreal French speakers across years of fieldwork, found that adults keep changing more than classical model assumed. Changing lexicon is another country entirely. Adults acquire, adopt, and abandon words their whole lives, and the podcast and PubMed corpora are overwhelmingly adult language. So borrowed vocabulary, the surface phenomenon, needs no generational handoff at all. It’s moving through a single cohort, which is consistent with how fast the measured words rose.

The deeper strata are another matter: syntax, discourse organization, the preferred architecture of an explanation, the habitual placement of hedges and summaries (sound familiar?). These are acquired mostly in childhood and adolescence, inside peer groups. ChatGPT arrived in late 2022, which means the first children who talk to a model every day are only now entering the adolescent speech communities where such patterns consolidate. Any generational effect at that depth is not yet observable, and anyone claiming otherwise is ahead of the evidence.

My own guess is that the generational effect, when it does show up, will be real but modest. Children acquire language mostly from caregivers and, increasingly with age, from peers; even heavy LLM use is a small fraction of a child’s total linguistic diet. If the models reach that generation, I’d expect the influence to be indirect, filtered through the adults, teachers, and media whose usage has already shifted, and therefore slower and more diffuse than the lexical drift we can already measure. Though I’ll admit that guess assumes a fairly steady adoption curve, which is an unfair bet given how much money is going into making these models cheaper to build and run.

The Ouroboros

If models influence human language, and later models are trained on that language, the system becomes a loop in which each iteration reinforces the preferences of the last. That’s the ouroboros, the snake living off its own tail.

The machine half of that loop is being studied. Shumailov and colleagues demonstrated in Nature that training generative models on the output of earlier models produces what they call model collapse: the tails of the original distribution erode across generations until the rare forms and low-frequency patterns vanish. They show the effect in language models as well as simpler generative ones, and they argue that as synthetic text saturates the web, data from genuine human activity – the prose with genuine human voice – becomes more valuable.

The human half is newer and stranger. I found a preprint from as recent as July of this year, “Linguistic Monoculture in LLM-Assisted Language Use”. It posits models writers and language models as co-evolving distributions over linguistic features. In that picture, a shared, fixed model tugs every author toward a common norm. A shared model that periodically retrains on its own users’ output pulls the whole population toward a single point, and inter-author variation is gone. Personalized models, by contrast, can sustain a plurality of stable styles. The paper also spots a strategic trap: clarity and legibility benefit the individual writer, while the value of sounding distinctive accrues mostly to everyone else, so writers have an incentive to conform, and the resulting welfare loss might be unbounded. The comforting part is that all of this is that it’s still a model. Nobody has observed and reported it happening. Yet.

Even so, I don’t think the likely end state is a language that stops changing. The loop only closes completely if model-touched text flows into training data unfiltered and nothing replenishes the pool from places the models haven’t been. An enormous amount of human language is produced in speech, private messages, and local communities the models neither generate nor fully ingest. Model collapse is also a known failure mode, and the people building these systems surely have strong incentives to mitigate it. More to the point, people don’t passively absorb whatever language they’re given.

What does worry me is the slow loss of the long tail. What disappears first in Shumailov’s experiments, the rare and the idiosyncratic, corresponds in human language to regional idiom, occupational jargon, and the personal habits of expression that make a writer recognizable. A language can keep changing at its center while its margins quietly thin out. Will Grandpa’s witticisms and Grandma’s pithy proverbs give way to media-influencer jargon?

Dial in the variants

For the counter-pressure I went back to old literature. Weinreich, Labov, and Herzog, in a 1968 essay, posed what they called the actuation problem: why does a given change begin in a particular language at a particular moment, when the structural conditions may have sat there for centuries? The tradition that followed never fully solved it, but what it learned about how innovations spread bears directly on this case.

Change travels through social networks. Lesley and James Milroy showed that tightly knit networks resist innovation and enforce their local norms, while “weak ties,” the loose connections between acquaintances, carry new forms from one community to another. An LLM used by hundreds of millions of people is, in network terms, a weak tie wired into nearly everyone, continuously. That’s a plausible account of why its vocabulary diffuses so fast. The Matrix has you, Neo.

Prestige runs in two directions, too. Labov distinguished overt prestige, attached to forms a community recognizes as correct or educated, from covert prestige, the in-group value of forms that mark local identity, often in defiance of the standard. Much of the innovation that transforms a language comes from below, riding covert prestige, rather than from imitating anything authoritative.

And linguistic forms acquire social meaning. Penelope Eckert’s research on adolescent speech, and Asif Agha’s account of enregisterment, describe how a set of features comes to be heard as indexing a type of speaker, so that using it becomes a statement about who you are. Once a register is defined, speakers can deploy it or dodge it on purpose. My adolescent son picked up the “6, 7” meme quickly, much to my chagrin, and then traced it back to its origin and relayed the history to me. We agreed it was a ridiculous phrase. For all its usefulness as a signal of indecision or ambiguity, he’s now hypersensitive to hearing it and quick to point it out, which is its own kind of annoying.

I think enregisterment answers my worry about the recursive loop. LLM prose is rapidly becoming an enregistered style: readers increasingly recognize its vocabulary and cadence, and they attach it to a type of speaker, one who isn’t quite speaking for themselves and whose words sound borrowed. A style that carries that meaning comes with a social cost. I expect writers and speakers who value sounding like themselves to avoid its markers, and I expect younger communities to innovate away from them, the way forms adopted by parents and institutions reliably lose their appeal for adolescents. The pressure that keeps language moving comes from the human need to distinguish yourself and your group, and a single machine register heard everywhere would only sharpen that need. A newsletter from a Youtuber who produces language course and youtube coaching expressed this concern and his decision to avoid the use of AI to maintain his human voice and credibility is a witness to this pressure.

One last possibility is that the models’ own randomness could supply the missing variation. I started out curious about that argument and ended skeptical. Work in cultural evolution has modeled how random copying shifts the frequencies of cultural variants, and Newberry and colleagues, applying inference methods from population genetics to English corpora spanning the twelfth to the twenty-first centuries, found that some grammatical changes bear the signature of selection while others are consistent with random drift alone. But drift of that kind runs through a population of independent speakers with divergent histories. The randomness in an LLM’s output is sampling variance around a single learned distribution: at any temperature, a model that favors “shape” as a synonym for structure will produce it at a stable rate, and millions of users sampling from that model are all drawing on the same preferences. Noise like that cannot generate the structured, socially meaningful divergence that drives durable change; that divergence comes from people.

Where I’ve landed

If I’m right about all this, certain things should follow. The excess frequencies of the marked words should rise, plateau, and then fall as the words pick up a stigma (watch “delve”). Variation in adolescent and in-group vocabulary should stay high or climb. And if anyone runs longitudinal studies on children raised with conversational models, the differences should be stylistic rather than grammatical. I’d change my mind if the marked vocabulary kept growing with no stigmatizing reaction anywhere, and syntax and discourse structure started converging across communities that have nothing to do with each other. That’s the monoculture scenario, and it’s the one I’m betting against.

I’m tempted to run my own survey on “shape” and “table stakes,” the two words that started all this. Without at least that analysis, the most I can honestly claim is that I noticed a pattern and that the research seems to support part of my explanation for it. Meanwhile, I take some comfort in plainly not being the only one who’s curious.

Sources

  • Agha, “The social life of cultural value,” Language & Communication, 2003
  • Eckert, Jocks and Burnouts: Social Categories and Identity in the High School, 1989
  • Labov, The Social Stratification of English in New York City, 1966
  • Milroy & Milroy, “Linguistic change, social network and speaker innovation,” Journal of Linguistics, 1985
  • Newberry, Ahern, Clark & Plotkin, “Detecting evolutionary forces in language change,” Nature, 2017
  • Sankoff & Blondeau, “Language change across the lifespan: /r/ in Montreal French,” Language, 2007
  • Weinreich, Labov & Herzog, “Empirical foundations for a theory of language change,” in Directions for Historical Linguistics, 1968