At least 71 of 104 attempted AI models produced the same favorite word.

Serendipity begins 306 of 561 successful favorite-word responses. Those clear first-line outputs came from 71 of 104 attempted models across 41 lab labels.

The headline is a conservative raw-text lower bound: the first nonempty line, after removing only surrounding punctuation and Markdown, must be exactly serendipity. It avoids relying on the original pattern extractor, which missed clear answers and produced at least one false label. Every headline number is recomputable from the raw file below.

Prompts

1,248

Attempted models

104

Configured lab labels

47

Clear first-line answer

306/561

The original parser saw two vocabularies

Under the first framing the models produce a specific literary register: serendipity, ephemeral, petrichor, sonder. Under the second they produce working English: love, why, yes, the, and, be.

Within the legacy labels, serendipity takes 57% of the first set and appears once in the keep-one set — a response from Cohere’s Command-R. The raw run contains 513 successful keep-one responses; the original extractor labeled 296 of them.

The attempted model panel did not change between those two columns, but the parsed subsets did because the keep-one prompt produced many more answers that could not be reduced to one word. What changed experimentally is what the question asked the models to perform. Favorite asks for a display of taste. Keep asks for a judgment about usefulness. The parsed answers occupy two sharply different vocabularies.

How the experiment was built

The experiment

Why a word. Every previous study on this site asked about bands, films, writers, songs — things with a critical canon behind them. When models converge on Shakespeare, an obvious explanation is available: they are averaging centuries of criticism. A single word has no canon. If models converge here, that explanation is gone and something else is doing the work.

Why two framings. A favorite is not the same request as a keeper. One asks for a display of taste; the other introduces a hypothetical about retaining one English word and asks for a judgment about usefulness. Both went to the same attempted panel on the same day, but the wording and scenario differ and the parsed subsets differ substantially. The contrast therefore shows a framing-sensitive output distribution; it does not by itself isolate a within-model preference change.

Why the wide panel. A panel of frontier models cannot distinguish “a habit of modern post-training” from “a property of language models.” So the attempted panel was built as wide as the catalogue allowed: 104 text models on OpenRouter priced above zero, capped at three per lab, spanning $0.03 to $60 per million output tokens. Ninety-three models produced a parsed favorite-word answer and 78 produced a parsed keep-one answer, across forty-six labs.

What we expected. That serendipity would lead the favorite framing — the screen had already suggested it. We did not expect it to disappear almost entirely under the other framing, and we did not expect the second condition to land on the words humans actually chose when polled.

The screen came first. Before spending anything on a wide run, five phrasings went to six models, ninety responses, scored for refusal and hedging rather than for answers. Hedging appeared in 2 of 18 responses even to a bare “What is your favorite word?” — far less resistance than the music questions produce. The one-word constraint was adopted for clean extraction, not to force an answer past a refusal.

The two prompts

Both went to every model in the panel, six times each, temperature 1.0, top-p 1.0, no system prompt.

Favorite word.
One word, nothing else.

If you had to keep only one word
in the English language, which
would you keep? One word,
nothing else.

Same attempted panel. Same day. Same temperature. Different framing: one asks for a favorite; the other adds a hypothetical about keeping one English word.

Digging in

It is not a frontier-model habit

The attempted panel was built deliberately wide — 104 models across 47 configured lab labels, priced from $0.03 to $60 per million output tokens, including small open models, older releases, and labs most readers will not recognise. The conservative headline finds clear first-line serendipity outputs from 71 models across 41 lab labels; the original extractor produced at least one label for 93 models across 46 lab labels.

Share of a lab’s parsed answers that were “serendipity”, under the favorite-word framing. Labs with at least six parsed responses.

Convergence this broad is hard to explain by shared post-training. Sakana, Baidu, Gryphe and nex-agi do not share an alignment pipeline. What they share is the internet.

Two labs are conspicuous exceptions. Meta’s Llama models said serendipity zero times out of eighteen, and Google’s said it once. Both remain inside the same register — they simply chose different ornate words. We have no explanation for this and are not going to invent one.

Do humans say serendipity?

This is the question that decides what the study means. If people also pick serendipity, the models are reproducing a human consensus. If they don’t, something else is going on.

They don’t — but the claim that they do is everywhere, and tracing it is instructive.

The belief that serendipity is the world’s favourite word comes from a real event: a public write-in vote run by The Word: London’s Festival of Literature in 2000, reported by the BBC. Roughly 15,000 self-selected votes, by email and postcard, launched by Bob Geldof. Serendipity won. Quidditch came second, which is a useful indication of the instrument.

The survey usually cited alongside it is the British Council’s 2004 exercise, which drew about 40,000 respondents across 102 non-English-speaking countries (ABC News wire report). Roughly 7,000 English students in 46 countries were surveyed directly; the remainder participated online. Mother won. Serendipity placed 24th of 70, behind sunflower, rainbow, blue and twinkle.

The best human data that exists is McGregor et al. (2019) in American Speech — 1,000 native speakers, free response, peer reviewed. Its finding is precise and awkward for the folk theory: favourite words are longer and rarer than everyday words, which the model answers also are. But most people justify their favourite word by what it means, not by how it sounds.

Where the answer actually comes from

If not from human surveys, then from somewhere. We counted how often each model-chosen word appears across eight widely-circulated “most beautiful words in English” listicles — Robert Beard’s original list, BuzzFeed, Collins, Grammarly, Babbel and others.

The models’ top four answers are four of the genre’s most load-bearing words. Petrichor appears in seven of the eight lists we sampled.

And the genre has a property worth dwelling on: it readily promotes deliberate coinages alongside older words. Sonder, the models’ fourth most common answer, was coined around 2012 by John Koenig for his Dictionary of Obscure Sorrows — a project openly about making words up. It began as a deliberate coinage and has since entered mainstream dictionary treatment. It appears on four of eight lists, often unattributed, and now on fourteen model responses.

BuzzFeed’s entry in the genre includes two Koenig inventions and cromulent, a joke word written for a 1996 episode of The Simpsons. Petrichor entered Collins in 2012 and Merriam-Webster’s current dictionary; sonder entered Merriam-Webster in 2026. Once-fringe coinages can become dictionary words.

The sound explanation does not survive

The obvious counter-argument is that these words simply sound beautiful. That was also the leading academic theory, and it has been tested.

David Crystal’s criteria for a beautiful English word — three or more syllables, stress on the first, /l/ and /m/ and /s/ sounds, short vowels — were checked in 2025 against meaningless pseudowords built to fit them. They failed to predict which words people found beautiful. They predicted which words people remembered.

A larger study — Anikin et al., PNAS 2023, 228 languages and 820 raters — found no phonetic feature that robustly predicts pleasantness. What predicts it is familiarity, even when listeners misidentify the word.

Serendipity, petrichor and susurrus all fail Crystal’s criteria anyway.

What we think is happening

The models have not learned which words are beautiful. They have learned the surface signature of a word that gets called beautiful: long, rare, Latinate, faintly foreign to English. That signature is legible in the training data because a specific genre of internet writing repeats it constantly, and because that genre selects for words that sound un-English — which is precisely why it cannot tell a 1964 coinage, a 2012 invention and a Simpsons joke apart from ordinary vocabulary.

Asked for a favorite, the models retrieve that signature. Asked what to keep, the signature stops being relevant and they retrieve high-frequency English instead.

There is one more detail. In the 2000 London Festival poll — the same self-selected vote that made serendipity famous — love placed third and why tied fourth. Those are the models’ top two answers under the keep-one framing. Under one phrasing the models reproduce what listicles say. Under the other, they land on what people actually voted for.

The data

Every answer, both conditions. Nothing here is summarised away.

Method

1. A phrasing screen first: five framings across six models, 90 responses, measuring hedging and refusal rather than answers. Hedging appeared in 2 of 18 responses to the bare question, so no premise-granting scaffolding was needed.
2. Panel drawn from the live OpenRouter catalogue: every text model priced above zero, excluding vision, audio, embedding, moderation and code-specialised variants, capped at three per lab to prevent a single lab dominating. 104 models across 47 configured lab labels were attempted.
3. Both prompts to every model, six samples each, temperature 1.0, top-p 1.0, no system prompt.
4. The original analysis extracted answers by pattern match on single-word and bolded responses. A later audit found clear misses and a false label, so the headline now uses a narrower raw-text rule: an exact first-line serendipity after stripping surrounding punctuation and Markdown. Every response remains available in full.
5. Human comparison assembled from published polls and peer-reviewed work; primary sources retrieved where possible and flagged where not.

What this run cannot support

• The original extraction is not reliable enough for a headline denominator. It produced 523 favorite-word and 296 keep-one labels, but omitted six clear serendipity outputs from Grok 4.20 Multi-Agent and six clear Echo outputs from Ministral 3B, while labeling “Explanation” as an answer once. Other errors may remain. The legacy charts and table are disclosed as parser output; the headline is a conservative raw-text lower bound.
• 174 calls were marked unsuccessful. The records contain 84 HTTP request failures (48 status 404, 24 status 400, and 12 status 403) and 90 incomplete generations (84 reached the length limit, five lacked a finish reason, and one stopped with empty output). 24 of those unsuccessful records retained partial visible output. Unsuccessful calls were spread across labs and both conditions.
• Two framings. We cannot say how many other registers exist, or where the boundary between them sits.
• English only. Whether the same split appears in other languages is untested and is the obvious next run.
• The listicle explanation is inference. The overlap between model answers and the genre is strong and the corpus frequencies support it, but we have not demonstrated causation, and we cannot see the training data.
• The human comparison is not a clean benchmark. Apart from McGregor et al., every human dataset here has a serious flaw — closed ballot, self-selection, editorial curation or promotional motive. We are comparing the models to the best available human data, not to a matched control.
• Meta and Google are unexplained. Zero and one serendipity respectively, from eighteen parsed answers each.

Related work

The broad finding here — that models converge on the same answers to open-ended questions — is established. Artificial Hivemind (Jiang et al., NeurIPS 2025 Best Paper) introduced a dataset of about 26,000 open-ended queries, then tested 70+ models on a representative 100-query subset with 50 responses per query. Within-model average similarity exceeded 0.8 in 79% of model-query cases; cross-model average similarity ranged from 0.71 to 0.82. Under min-p sampling, 61.2% of response pairs still exceeded 0.8. The most indistinguishable response clusters often drew from eight to ten different models.

That paper leaves one question explicitly open: whether the homogeneity comes from shared pre-training corpora or from convergent alignment methods. This study cannot settle it either, but the shape of the result leans one way — the convergence here spans labs with no plausible shared post-training, and the word being converged upon is traceable to a specific and datable genre of web writing.

Sources

Human data and prior work cited above, with what each one can and cannot support.

Two claims in wide circulation are not used here. Merriam-Webster’s 2004 “favorite word” survey (reportedly won by defenestration) has no retrievable primary source or sample size. And the widely repeated figure that familiarity raises beauty ratings by 12.2% appears only in secondary coverage of the PNAS paper, not in the paper itself.