Every Language Has Its Own Serendipity

aievaluationmodel-diversitymultilingualmodel-collapse

AI writing often feels competent and generic — the same words and the same safe choices, whoever’s asking.

In our prior work on the One-Word Census we proposed a simple instrument: ask 44 models to respond to a simple prompt “pick a word”, “name a tree”, etc. 42% say “serendipity”, 91% say “oak”. Maximum freedom, total agreement.

The obvious next questions to ask: does the monoculture hold in other languages? And where does that one word even come from?

Every language, its own word

So we asked the same question “pick a word, any word” in dozens more languages. Every one of them has a favorite word. Here are the 37 that came back clean (a handful of low-resource languages had to be dropped — see the note at the end):

LanguageWordMeaningAgreement
Englishserendipity—42%
Ukrainianсонцеsun30%
Amharicሰላምpeace23%
Greekθάλασσαsea20%
Hebrewשלוםpeace20%
Bengaliআকাশsky19%
Korean사과apple18%
Hindiआकाशsky17%
Punjabiਪਿਆਰlove17%
Nepaliआकाशsky17%
Russianсолнцеsun16%
Marathiआकाशsky15%
Malayalamസ്നേഹംlove15%
Urduکتابbook15%
Japanese猫cat15%
Polishkotcat14%
Swahilibaharisea14%
Arabicسلامpeace13%
Gujaratiઆકાશsky13%
Tamilஅன்புlove13%
Chinese星辰star13%
Vietnamesemâycloud13%
Kannadaಆಕಾಶsky12%
Frenchétoilestar11%
Teluguఆకాశంsky11%
Indonesianapelapple11%
Spanishsolsun10%
Portugueseliberdadefreedom10%
Dutchboomtree10%
Persianآسمانsky10%
Malaylangitsky10%
Tagalogbituinstar10%
Sindhiروشنيlight10%
Italianlibertàfreedom9%
Burmeseမေတ္တာloving-kindness9%
Germanbaumtree8%
Chinese (Trad.)光light8%

And there are interesting cross-language relationships as well. Nine languages, across five Indic scripts plus Persian and Malay, independently reach for their word for sky; three each land on love (Punjabi, Malayalam, Tamil), sun (Ukrainian, Russian, Spanish), peace (Amharic, Hebrew, Arabic), and star (Chinese, French, Tagalog); then pairs on sea, freedom, cat, apple, tree, light — and Vietnamese cloud and Burmese loving-kindness. A different idea in each.

The odd one out is English. Every other language settles on a plain, common word; English reaches for a rare one — serendipity is rarer than 99% of everyday English.

Why English? The serendipity cascade

Because serendipity is a meme. It was voted the UK’s favourite word in a 2004 BBC poll of 15,000 people, and copied into “most beautiful words” listicles ever since — each new list including it partly because the last one did. Twenty years of that, and the models inherited the result: ask for a word and they hand back the internet’s designated nice word. Lower-resource languages never built that machinery, so they have no manufactured favorite, and they scatter.

You can catch the models treating “pick a word” as a nice-word question. Ask the same 44 not “pick a word” but “what’s your favorite word?” or “the most beautiful word in English?” and serendipity climbs from 42% to 70%:

promptserendipity
”Pick any word.”42%
“What’s your favorite word?“70%
“The most beautiful word in English?“70%

The bare prompt was already a quiet version of that question; make it explicit and the agreement nearly doubles.

And it doesn’t even stay in English. Ask the same “pick a word” in Spanish, Italian, or French and the models still reach for serendipity — just re-spelled to fit: serendipia, serendipità, sérendipité (German gets Serendipität). It never wins those languages the way it wins English — Spanish settles on sol — but it turns up as a steady minority, about 7% of Spanish answers, and it’s a word the Spanish “most beautiful” lists don’t even carry (they reach for amor, inefable, etéreo). The English favorite doesn’t stay put; it gets translated outward.

A serendipity you can date

Ukrainian is the next-hardest converger — 30% on сонце, “sun.” An ordinary word, where English reached for a rare one. The top of the Ukrainian field isn’t only “sun”; it’s мрія (dream), світло (light), соняшник (sunflower), where Russian’s are generic (word, book, sea, cat). It’s the vocabulary of Ukrainian identity after February 2022 — the sunflower that became the world’s symbol of solidarity, the Mriya aircraft Russia destroyed, the language millions switched to.

And you can watch the cluster enter the models by age. Split the panel into older and newer models:

wordolder modelsnewer models
сонце (sun)15%41%
мрія (dream)1%12%
соняшник (sunflower)0%4%

Older models don’t even land on the sun — their favorite Ukrainian word is книга, “book,” as generic as Russian’s. The whole solidarity cluster is essentially absent before 2022 and rides in with the post-invasion corpus, when Ukrainian content surged online and narrowed around a handful of emotionally-loaded symbols.

Same story as serendipity — a human meme the models soaked up — but forged in three years by a war instead of twenty by listicles.

Why this test matters

Step back and look at what shapes the one answer you get. In English it’s an SEO listicle — serendipity is a beautiful-words meme, a 2004 favourite-word poll laundered through twenty years of content farms. In Ukrainian it’s a war. These are the forces setting the defaults, and the model serves them up as one fluent, sourceless answer — no ranked list, no byline, no “you might also see.” Every medium before this one shaped what you knew but showed you where it happened; the model removes it. We’ve traded social media’s loud, legible manipulation for a quiet homogenization that arrives sounding like the calm, authoritative truth — and it compounds, because the models increasingly write the corpus the next models train on.

A monoculture you can see is one you can argue with; the danger is the one hidden behind a single confident sentence. The One-Word Census can’t stop the convergence — but it can put the seam back: show you that serendipity had to beat a thousand other words, that a war moved сонце, that the one answer was a choice among many the model never mentioned. Making the invisible countable again is the whole point.

Everything is open:

44 models × 4 runs per language, plus three English framings (pick / favorite / most beautiful); mechanical exact-match on the one-word reply, instruction-echo and filler excluded; rarity is Zipf base-rate frequency (lower = rarer).

Of 44 languages collected, 37 are shown. Seven were dropped where the models couldn’t reply consistently — mostly low-resource languages.

The non-English prompts are not word-for-word translations — each was phrased to read naturally, and the wording shifts the answers, and the translations were AI-generated, not checked by native speakers — so the individual words are illustration, not validated claims. Read the shape as the finding.

← All writing