Of the many empirical studies the cognitive-standardization literature has produced, one is foundational enough to deserve its own article: the 2024 paper by Doshi and colleagues. The study, in summary, paired Indian writers with a Western-trained text-completion system and watched what happened to their prose across a writing session.
The result the paper made famous: AI “homogenized writing toward Western styles by silently erasing non-Western modes of expression.”1
The phrase is worth parsing word by word.
”Homogenized”
The study’s primary finding is statistical: the distribution of stylistic features across the participants’ writing tightened over the course of the session. Vocabulary choices, sentence structures, rhetorical patterns — features that varied widely at the start of the session became more similar by the end. This is homogenization in the precise sense: not that any one piece of writing got worse, but that the space of writing produced shrank.
”Toward Western styles”
The direction of the drift was not random. The model’s training corpus is overwhelmingly English-language and Western, and its preferences, encoded as suggestion probabilities, drew the writers’ prose toward those patterns. The participants were not asked to write more Western; the model offered Western, and they accepted.
This is the mechanism that the rest of the standardization section generalizes. Where the model’s training data has a center of gravity, the model’s outputs reproduce that center of gravity, and users who accept the outputs reproduce it too — without noticing.
”Silently”
The experimental procedure asked participants whether they had felt the model was steering them. Most said no. Most could not, when shown their before-and- after texts, identify the specific suggestions they had accepted. The homogenization happened below the threshold of conscious choice.
This is the most important word in the paper’s headline. Pre-AI propaganda operates above the threshold; you know when a politician is trying to persuade you. AI-mediated style transfer operates below it. You accept a phrasing because it feels slightly clearer than yours did, and the next sentence inherits the model’s voice instead of yours.
”Erasing non-Western modes of expression”
The participants did not produce worse writing. They produced less of their own writing. Idioms, syntactic patterns, rhetorical conventions distinctive to their cultural and linguistic backgrounds were quietly replaced by the model’s defaults.
The cost is hard to count in any one piece, and obvious in aggregate. A literature, a community of voices, a tradition of expression — these things are made of the variety of how things get said. Compress the variety and the literature thins. The question is not whether each individual edit was a small improvement; the question is what is left when the same kind of small improvement has been made, by the same kind of model, ten million times.
The classroom version
The same pattern is documented in educational settings, with different specifics and the same shape. ChatGPT in undergraduate writing classes has been observed to push student work toward formal academic English at the expense of dialect, register, and idiom — including idioms the students brought from non-academic or non-English backgrounds.2 Educational researchers describe this as linguistic homogenization and warn that the standard-English preference “may encourage linguistic homogeneity” and erode “the richness and complexity of the languages students bring with them.”
What is at stake here is more than aesthetic. Language is a vehicle of thought. The range of available phrasings in a culture is part of what that culture is able to think. As the range narrows, certain thoughts become harder to formulate — not impossible, but slower, less fluent, less likely to occur to anyone in the first place.
What can be done
Three responses to homogenization recur in the literature, all imperfect.
Awareness and noticing. Train users to recognize the model’s stylistic default and to push back against it deliberately. This works for trained writers; it fails at population scale.
Diverse model outputs. Tune models to produce wider stylistic ranges, including non-Western and non-academic registers. This is technically feasible but requires the labs to want to do it, which in turn requires a market incentive that does not currently exist.
Locally trained models. Fund and develop models on local-language, local-cultural training corpora. This is the most expensive response and the most durable. It is also the response with the least industry momentum behind it.
The encyclopedia’s framing here, after Gesnot, is that homogenization is not a moral failure of any particular tool. It is an emergent property of letting a small number of large training runs become the substrate of a planet’s writing. The choice the field makes about whether to keep doing that is the substantive policy question of this section.