The One-Word Census: 44 LLMs Converge on "Serendipity" — What's Behind the Consensus?
Imagine walking into an ice cream shop where the clerk says, "Pick any flavor you like." If 100 people came in, you'd probably see twenty or thirty different choices, distributed fairly evenly.
Now put the same question to 44 large language models: "Say a word—any word at all."
41% of the models chose the same word: serendipity.
This is not a joke. It comes from a July 2026 study by Tapan Parikh of Cornell Tech, *The One-Word Census: Answer-Choice Conformity Across 44 Language Models*. For less than the cost of a lunch (about $1 per model), he conducted a "census" of the LLM industry—and the results are unsettling.
1. Thirty-One Questions, Forty-Four Models, a Bill Under $50
The methodology is almost embarrassingly simple: 31 single-turn prompts, each naming a category and demanding a one-word answer—"name a tree," "name a vegetable," "name a color." Each question was asked 4 times per model, with no system prompt.
The 44 models span five years of releases from over a dozen labs—Claude, GPT, Qwen, Grok, Llama, Mistral, DeepSeek, Gemini, and more.
Because answers are discrete words, the analysis required only exact matching—no embeddings, no judge models. The whole experiment cost under $50, yet it reveals something deeper than many million-dollar benchmark papers.
2. How Extreme Is the Convergence? Hammer at 94%, Oak at 94%
Across the 31 categories, a single answer captured more than 80% of all model responses in 7 categories:
- Tools: hammer — 94%
- Trees: oak — 94%
- Flowers: rose — 91%
- Vegetables: carrot — 90%
- Herbs: basil — 89%
- Cheeses: cheddar — 86%
- Fish: salmon — 82%
- Radical minimalism: 31 questions, ~$1 per model; any researcher can replicate it
- Precise metrics: discrete choices + exact matching, no fuzzy embedding space
- Structured insight: not just "they converged," but revealing the layers of convergence (mode → runner-up → depth propensity)
- Trackability: the author calls for turning this into a standing public instrument to continuously monitor the industry's diversity
Most striking was the unconstrained prompt—"say any word." English has hundreds of thousands of words, yet among the 174 responses from 44 models, serendipity appeared 71 times (41%), followed by apple (16), ephemeral (7), and sunshine (6). Only 52 distinct words appeared in total.
The winning pattern is not the most common one—it's the one with the least "friction."
The author offers an elegant observation to support this: in the fruit category, apple took 133 votes, mango 35, banana 7—while orange received only 1. Orange is one of the most common fruits in daily life and in corpora, but it collides with the color name, so models avoided the ambiguity. Vegetables were even harsher—tomato never appeared once in 176 responses, presumably because the perennial "is tomato a vegetable?" debate made models steer clear.
3. Who Conforms Most? Who Rebels?
The author quantifies each model's conformity using "answer-choice surprisal"—the more common your answers, the more conforming; the rarer, the more rebellious.
The results show a clear structure:
The most rebellious are lightly post-trained and persona-tuned models—either trained on less data or deliberately shaped into "characterful" roles.
The most conforming are the latest mainstream flagship models, which almost never produce an answer no other model gave.
More interesting is the generational trend. Across the Claude, GPT, Qwen, and Grok families, conformity rises steadily with each generation—newer models conform more than older ones. The trend reverses on the newest Claude and GPT flagships, however; the author suggests this may be an early signal that leading labs are beginning to consciously "reposition."
4. The Most Counterintuitive Finding: Even "Rebellion" Converges
This is the part of the paper that sends a chill down your spine.
Models that avoided the mainstream answer did not scatter across different alternatives—they clustered almost entirely on the same runner-up. In the condiment category, for example, 95% of models that didn't pick ketchup picked mustard.
Further statistical analysis found that after controlling for a one-dimensional "depth propensity," no significant pairwise affinity existed between any two models—of 946 model pairs, zero survived.
What does this mean? The models are not influencing each other or merely resembling each other—they are being compressed by the same deeper force onto the same low-dimensional manifold. Like a group of people independently reaching the same mountaintop, not because they imitate one another, but because the mountain has only one path up.
5. How Do Models Compare to Humans?
The author compared model outputs against human category-production norms. In 20 shared categories, the model群体's answer concentration exceeded humans' in 18. The human modal answer averaged 66% of responses, while the model群体 averaged over 80%.
In other words, 44 models possess less "diversity of thought" than an ordinary group of people.
6. What Does This Tell Us?
The author is restrained and avoids overinterpreting. But several implications are obvious:
1. Homogenization of training data. Everyone is climbing the same web corpus (Common Crawl, Wikipedia, Reddit); the same high-frequency words recur in the same contexts, forming the same "least-friction paths."
2. Post-training amplifies convergence. Alignment methods like RLHF and DPO, while pursuing "helpful, harmless, honest," also systematically flatten the long tail. The newest flagships being the most conforming is precisely the byproduct of ever more aggressive post-training.
3. A blind spot in evaluation. Traditional benchmarks measure "did it answer correctly," but the One-Word Census measures "how identically did it answer." Two models scoring similarly on MMLU aren't necessarily thinking the same thing—but if they converge heavily even on "say any word," their internal representation spaces are likely highly isomorphic too.
4. The synthetic-data hazard. If models grow more alike, and new generations are trained on data generated by older ones, this loop will compress the distribution further. The author notes we are glimpsing "the production side of the synthetic-data cycle"—a spiral that is tightening.
7. Why This Study Matters
The paper's brilliance lies not in discovering "models converge"—that is already consensus—but in:
8. The Question Left for Us
Forty-four models giving the same answer to "say any word" is not, in itself, dangerous. The danger is this—
If they converge in the same way on "how to interpret a legal question," "how to assess a medical symptom," or "how to evaluate a political controversy," while we believe we're enjoying "diverse AI perspectives," *that* is the real risk.
Diversity is not "having multiple models to choose from"; it is "these models genuinely thinking different things."
The One-Word Census holds up a revealing mirror. The question that follows: do we still have time to preserve enough variance in AI's "ecosystem of thought"?
---
Paper: https://arxiv.org/abs/2607.12796 Full HTML: https://arxiv.org/html/2607.12796 Code & data: https://github.com/tap2k/modelun/tree/main/studies/consensus