Models write alike, not necessarily worse

The claim

Y Combinator’s requests for startups for its Winter 2027 batch include “LLMs That Write Well”. Its author, Jared Friedman, writes that as models improved at code and math they “have actually gotten worse at writing”, and that “the current models really do sound more robotic than the models from 2024 and 2025.” He calls it “a structural issue”: “The same reward signals that make models good at code actively flatten voice and style.” The major labs, he says, will always choose coding and agentic work over style, which leaves room for a startup that makes “a model that people like talking to.”

That is three claims: models write worse, code rewards are the cause, and the labs will not fix it. Papers from the last six months test the first two.

What has been measured is sameness

In August 2026, researchers at Duke compared 68 models from 12 providers, released between March 2023 and July 2026. They gave each model ten items from a standard creativity test, in which the task is to list uses for an everyday object, and 100 real open-ended user questions. They embedded the answers and measured how far apart different providers’ models sit on the same prompt. On the uses task the distance fell from about 0.50 to below 0.40. On the open questions it fell from about 0.34 to just above 0.32, mostly from 2025 on. All 1,000 resamples balanced across providers showed the decline.

The paper is labelled preliminary. It measures similarity, not quality, and it says it cannot separate distillation or shared training data from other causes. An earlier study found the same gap from the other side: across a wide set of models, answers to creativity tests were far more similar to each other than people’s answers are to each other (Wenger and Kenett).

So “more alike” is measured. “Worse” and “robotic” are not. The Duke study does not measure how readers judge the voice.

The flattening comes early, and the judges like it

Researchers at the University of Sheffield traced one open model family, Olmo 3, through its training. The “Think” model lost 62% of the base model’s answer diversity in the first supervised stage, 4% more in preference tuning, and regained 4 points in the reinforcement-learning stage. The “Instruct” model lost 38%, then 23%, then 5%. Reinforcement learning applied straight to the base model kept a median 94%.

The Think supervised stage trained on completions from two teacher models, and the authors read the loss as copying a narrow source. Suppressing the model’s chain of thought left diversity unchanged, so the narrowing sits in the weights, not in how the answer is written. Their conclusion: it is set by the training data and cannot be fixed at inference alone.

This partly supports the YC mechanism. Among the models trained with reinforcement learning only, pass-or-fail code rewards narrowed answers most (88% of base diversity kept, against 99% for instruction-following rewards). But in the pipelines that resemble real ones, the reinforcement step moved little. Frontier labs’ pipelines are not public.

A second study isolated post-training on fiction. Using four checkpoints of one 32-billion-parameter model, it found each stage compressing the writing. The share of surprise and conflict in New Yorker stories fell from about 41% to about 20%, and style differences between story types dropped by more than 90%. The better the human baseline, the bigger the gap: professional literary fiction was compressed most.

Where does the pressure come from? One answer is the preference data. Annotators favour familiar text, a well-known effect in psychology, and Zhang and colleagues show this typicality bias helps cause mode collapse. A prompt that asks the model to list several answers with their probabilities raised creative diversity 1.6 to 2.1 times, with no training, and more capable models gained more. So prompting recovers part of what training removed.

The same Olmo study asked a model judge to compare post-trained stories with the base model’s. The Think and Instruct models, each down 63% in diversity on creative writing, won more than 97% of the comparisons. The judge calls the flatter output better, and the preference data that helped flatten it shares that taste. A judge is one voter: on New Yorker captions a strong model agreed with the crowd about as often as one random voter did (A model judge is one voter).

Whose ear?

Chakrabarty and colleagues ran a preregistered test of writing in the style of 50 award-winning authors. Readers trained in a creative-writing MFA strongly disfavoured prompted AI text, for fidelity (odds ratio 0.16) and quality (0.13). College-educated general readers did not mind: no fidelity preference, and a preference for the AI on quality (1.82). “Robotic” is what trained ears hear.

The same study shows what moves it. Fine-tuning a model on one author’s complete works reversed the verdict. MFA readers then favoured the AI for fidelity (8.16) and quality (1.87), general readers more strongly (16.65 and 5.42), and detectors flagged 3% of the fine-tuned text against 97% of the prompted text. The cost was $25 to $276 per author, a median of $81. Fine-tuning used each author’s complete works, so this is not a test of context against weights.

What holds up

The first claim holds in a narrower form: models are becoming more alike, with the clearest effect on creative tasks. The second holds weakly: code rewards narrow answers, but supervised data and preference data narrow them first and more. The third is a forecast. What fixed voice in the one study that tested readers was one author’s own writing, not a better single style.

What it means for our work

Our subject is ideas, not prose, and ideas converge too. Thirty answers to one skyscraper brief from one model gave 26 named for the tide (Forty concepts, three ideas). If sameness comes in with the training data and the preference data, a better prompt will not remove it, and the input has to come from outside the model.

That input is a person. A framework is a written account of how one person reasons, and a model that carries it gave answers a judge could trace to their author 88% of the time against 25% by chance (A way of thinking travels). The author study points the same way for voice: what moved the model was one person’s material, not a better global style.

The other half is choosing. If taste were shared, a flatter and safer text would be the right answer for everyone, and the model judge’s verdict would settle it. About 44% of what decides a rating is one person’s own stable taste, and 23% is shared (Taste is mostly personal). A model people like talking to is a model that knows which people.