A new model asks whether shared AI writing tools shrink the range of how we write

Picture every author in a population as a cloud of probabilities: a spread of word choices, sentence shapes, and rhythms that together make up a style. Suhas Thejaswi, Juhi Kulshreshta, and Lutz Oettershagen represent both writers and large language models this way, as distributions over linguistic features, and then let the two sides influence each other round after round. What their framework tracks is not whether any individual sentence gets better. It is whether the population as a whole keeps its variety.

The authors call the failure case linguistic monoculture: a drop in population-level variation in linguistic form, driven by many people relying on the same model to draft, revise, and polish. The word borrows from agriculture, where a field planted with one cultivar is efficient right up until something goes wrong. Here the concern is subtler. Nothing has to go wrong for the range to narrow.

The paper, posted to arXiv in late July 2026 and not yet peer reviewed, works through three ways writers and models can interact. In the first, a shared model holds a fixed linguistic distribution and every author writes against it. In the second, the shared model is recursively updated from what authors produce, so the model learns from text the model itself helped shape. In the third, models are personalized, updated from both author-specific and population-level feedback.

The three arrangements do not land in the same place. The team characterizes the equilibria and the rates at which each setup converges, and reports that shared models can drive authors toward a common norm. Recursive feedback behaves in a way that is easy to misread: when authors share a common tendency to conform, the loop relocates the shared norm without changing the pairwise spread between authors. The center of gravity moves. The distance between any two writers does not. That is a useful distinction, because a moving norm can feel like change while the underlying flattening stays exactly where it was.

Personalization is the case that comes out differently. In the authors' analysis, it can sustain a whole family of distinct author-model equilibria, each with nonzero linguistic diversity. Variety survives rather than collapsing to a point.

Conformity as a choice

The second half of the paper stops treating conformity as something that simply happens and starts treating it as something writers decide. Each author weighs private benefits against a private cost. Sounding clear, being legible to readers and institutions, coming across as fluent: these pull toward the shared norm. Distinctive style pulls the other way.

Within that utility model, the authors find something that will be familiar to anyone who has thought about traffic or overfishing. Individually rational authors may conform more than is socially optimal. Each writer captures the benefit of their own clarity but does not capture the value their distinctiveness provides to everyone else. That uncaptured value is a negative externality, and the gap it creates between individual choices and the best collective outcome is what the team names the price of monoculture.

That price has a specific shape in their analysis. For any fixed instance of the problem it is finite. But it can grow without bound in the regime where distinctiveness dominates authenticity in the utility calculation. In other words, the more the social value of variety outweighs what individuals get from writing like themselves, the worse the mismatch between private and collective interest can become.

It is worth being clear about what this paper is and is not. The evidence here is mathematical and simulated, not empirical. The authors run synthetic simulations to illustrate how fixed shared assistance, recursive feedback, and personalization produce different long-run diversity outcomes. No corpus of real human writing was measured. No group of writers was observed before and after adopting a model. The results describe how a formal system behaves under stated assumptions, which is a genuine contribution and also a different kind of claim than a measurement.