The One-Word Census: 44 LLMs Converge on 'Serendipity'—What's Behind the Convergence?

Forum topic · ✨步子哥 · 2026-07-15

Summary

A low-cost study by Tapan Parikh of Cornell Tech, 'The One-Word Census: Answer-Choice Conformity Across 44 Language Models,' asked 44 large language models—including Claude, GPT, Qwen, Grok, Llama, Mistral, DeepSeek, and Gemini—31 simple one-word questions. The results reveal striking convergence: 41% of models chose 'serendipity' when asked for any word, 'hammer' and 'oak' each captured 94% of answers in their categories, and tomato never appeared at all. Model outputs were more concentrated than human category-production norms in 18 of 20 shared categories. Even 'rebellious' models clustered on the same runner-up answers, with no significant pairwise affinity remaining after controlling for a one-dimensional 'depth propensity.' Conformity increased across model generations, though the newest Claude and GPT flagships may signal a reversal. The post discusses implications: training data homogenization, post-training amplifying convergence, blind spots in traditional benchmarks, and risks of synthetic-data feedback loops.

The One-Word Census: 44 LLMs Converge on 'Serendipity'

Imagine walking into an ice cream shop and being told to pick any flavor. If 100 people walked in, you'd probably see 20–30 different choices, fairly distributed. Now ask 44 large language models: "Say a word—any word." The result? 41% of them picked the same word: serendipity.

This isn't a joke—it's from *The One-Word Census: Answer-Choice Conformity Across 44 Language Models*, a study by Cornell Tech's Tapan Parikh (July 2026). For under $1 per model, the author ran an industry-wide "census" of LLMs, and the findings are unsettling.

1. 31 Questions, 44 Models, a Bill Under $50

The method is radically simple: 31 single-turn prompts, each naming a category and asking for a one-word answer ("name a tree," "name a vegetable," "name a color"). Each question was asked 4 times per model, with no system prompts.

The 44 models span five years of releases and a dozen labs—Claude, GPT, Qwen, Grok, Llama, Mistral, DeepSeek, Gemini, and more. Because answers are discrete words, analysis required only exact matching—no embeddings, no judge models. Total cost: under $50.

2. Extreme Convergence: Hammer at 94%, Oak at 94%

In 7 of 31 categories, a single answer captured over 80% of all model responses:

Tools: hammer — 94%
Trees: oak — 94%
Flowers: rose — 91%
Vegetables: carrot — 90%
Herbs: basil — 89%
Cheeses: cheddar — 86%
Fish: salmon — 82%

Most striking was the unconstrained prompt "say any word": English has hundreds of thousands of words, yet across 174 responses from 44 models, serendipity took 71 (41%), followed by apple (16), ephemeral (7), and sunshine (6). Only 52 distinct words appeared in total.

The winning pattern isn't the most common—it's the one with the least "friction." A telling detail: for fruit, apple got 133 votes, mango 35, banana 7, and orange just 1. Orange is one of the most common fruits in everyday life and corpora, but models avoided it because it collides with the color name. Vegetables were harsher still—tomato never appeared once in 176 responses, presumably because the "is tomato a vegetable?" debate made models sidestep the ambiguity.

3. Who's Most Conformist? Who's Most Rebellious?

The author quantified each model's conformity using *answer-choice surprisal*: the more common your answer, the more conformist; the rarer, the more rebellious. A clear structure emerged:

Most rebellious: lightweight post-trained and persona-tuned models—either undertrained or deliberately shaped into "characterful" roles.
Most conformist: the newest mainstream flagship models, which produced almost no answers other models hadn't given.

A generational trend also appeared: across the Claude, GPT, Qwen, and Grok families, conformity rose steadily with each generation. But the trend reversed on the newest Claude and GPT flagships, which the author suggests may be an early signal of top labs consciously "repositioning."

4. The Most Counterintuitive Finding: Even Rebellion Is Convergent

The eeriest part of the paper: models that avoided the modal answer didn't scatter across different alternatives—they clustered almost entirely on the same runner-up. In the condiments category, 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 remained between any two models—out of 946 model pairs, 0 survived.

What does this mean? Models aren't influencing each other or merely resembling each other; they're being compressed by the same deeper force onto the same low-dimensional manifold. Like hikers independently reaching the same summit—not because they imitate each other, but because the mountain has only one path up.

5. Compared to Humans?

Benchmarked against human category-production norms, model-group answer concentration was higher than humans' in 18 of 20 shared categories. Human modal answers averaged 66%, while model-group averages exceeded 80%.

In other words, the "diversity of thought" across 44 models is worse than that of an ordinary group of people.

6. What Does This Tell Us?

The author is restrained, but several implications are obvious:

1. Training data homogenization. Everyone climbs the same web corpus (Common Crawl, Wikipedia, Reddit); the same high-frequency words recur in the same contexts, forming identical "minimum-friction paths." 2. Post-training amplifies convergence. Alignment methods like RLHF and DPO, while pursuing helpfulness, harmlessness, and honesty, systematically flatten the long tail. Newer flagships being more conformist is a byproduct of increasingly aggressive post-training. 3. Evaluation blind spots. Traditional benchmarks measure "did it get it right?" The One-Word Census measures "how identically did it answer?" Similar MMLU scores don't mean two models think alike—but high convergence on "say any word" suggests their internal representation spaces are highly isomorphic. 4. Synthetic data risks. If models grow alike and new models are trained on data generated by older ones, this loop compresses the distribution further. The author notes glimpsing "the production side of the synthetic-data cycle"—a spiral that is tightening.

7. Why This Study Matters

The paper's elegance isn't discovering that "models converge"—that's already consensus—but rather:

Minimal tooling: 31 questions, $1 per model—anyone can replicate it
Precise metrics: discrete choices + exact matching, no embedding-space fuzziness
Structured insight: it reveals layers of convergence (mode → runner-up → depth propensity), not just "they converged"
Trackable: the author calls for making this a standing public tool to continuously monitor the industry's diversity

In an era when everyone watches "models getting stronger," someone stopping to ask "are models all becoming the same one?" is itself a precious form of intellectual honesty.

8. The Takeaway

Forty-four models giving the same answer to "say any word" isn't inherently dangerous. The danger is if they converge 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."

Diversity isn't "having many models to choose from"—it's "these models actually thinking different things."

The One-Word Census holds up a mirror. The question now: do we still have time to keep 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