The problem of homogeneity is definitely a serious epistimc challenge facing all of us. A new opinion paper published in Trends in Cognitive Sciences by Sourati, Ziabari, and Dehghani (2026) tackles this problem head-on. The paper discusses what happens when everyone uses the same handful of models to write, reason, and form opinions. The answer, according to these authors, is a kind of cognitive flattening that should concern anyone who cares about how people learn.
Sourati et al. (2026) synthesize evidence from linguistics, psychology, computer science, and cognitive science to argue that large language models are narrowing human diversity across three dimensions: language, perspective, and reasoning. The claim is specific. LLMs are trained to predict the most statistically probable next token in a sequence, which means they default to the most frequent patterns in their training data.
Rare expressions, minority voices, and culturally grounded ways of using language get smoothed over. What comes out is a polished average, and that average doesn’t trend toward some neutral middle. It trends toward the norms of English-speaking, Western, liberal, highly educated populations. WEIRD societies, in the technical sense.
I’ve covered similar ideas from the cognition side. Shaw and Nave (2026) wrote about cognitive surrender, the gradual offloading of reasoning to AI. Kosmyna et al. (2025) showed at the neural level that ChatGPT-assisted writing produces weaker brain engagement than independent writing. And Fan et al. (2025) documented how students skip metacognitive steps when AI is available. Sourati et al. are saying something broader: these aren’t isolated effects. They’re symptoms of a systemic homogenization that operates at the level of language itself.
The Language Problem Is Concrete
The linguistic evidence Sourati et al. (2026) present is striking. When LLMs polish writing, whether it’s a Reddit post, an academic abstract, or a personal essay, the resulting texts converge in complexity and style. The predictability of author characteristics drops. You can no longer tell from the writing whether the person is liberal or conservative, young or old, male or female. Adjusting the model’s temperature or using persona-based prompting doesn’t fix this. The homogenization persists.
How LLMs Flatten Perspectives and Shift Opinions
Sourati et al. (2026) also present evidence on perspective diversity. Using instruments like the World Values Survey and the Moral Foundations Questionnaire, researchers have found that LLM outputs show substantially less variance than human responses. The models align closely with WEIRD response patterns and show weak representation of non-WEIRD perspectives, even when explicitly prompted to simulate diverse viewpoints. Attempts to fix this through prompting or fine-tuning have produced surface-level variation at best, reproducing out-group stereotypes or collapsing to the socially “correct” mean.
The part that genuinely concerns me is the persuasion evidence. Sourati et al. (2026) cite studies showing that participants who cowrote with opinionated language models, ones engineered to frame social media positively or negatively, tended to mirror the model’s stance in their writing. And it didn’t stop there. They shifted their own opinions in subsequent attitude surveys. People adopted the model’s framing without realizing it. That’s not a hypothetical concern about influence. That’s a documented shift in how people think after interacting with a tool they believe is neutral.
For educators, this raises a question I don’t think we’ve grappled with seriously enough. If students use AI to help formulate arguments, draft essays, or prepare for debates, whose perspective are they actually articulating? The student’s, or the model’s? Niloy et al. (2024) found that ChatGPT can enhance creative writing on the surface, but at what cost if the “creativity” converges toward a single stylistic and perspectival norm?
The Reasoning Diversity Paradox
The reasoning section of this paper carries the most weight for anyone who cares about pedagogy. Sourati et al. (2026) point out that groups composed of individuals who reason differently consistently outperform groups of the highest-ability individuals. Diversity in reasoning is a collective asset. LLMs threaten that asset because they push everyone toward the same dominant reasoning style.
Chain-of-thought prompting, for example, is optimized for linear, explicit inference. That’s useful for many tasks, but it may also disincentivize abstract or intuitive approaches to problem-solving that don’t follow a step-by-step template. The authors cite a vehicle classification study where CoT prompting made GPT-4o four times slower to learn correct labels. The step-by-step reasoning overgeneralized from regular patterns and couldn’t handle contextual exceptions. That’s a concrete example of how optimizing for one style of reasoning creates blind spots.
In creative ideation experiments, participants who used LLMs generated a larger number of detailed ideas, but their outputs were judged to be semantically similar across participants. The models boosted individual volume and reduced collective diversity. I’ve been saying for a while that AI improves the product without always improving the process. Sourati et al. give that claim a structural explanation: when everyone reasons through the same model, the output converges no matter how different the individuals are.
The McDonaldization of Thought
Sourati et al. (2026) borrow sociologist George Ritzer’s “McDonaldization” concept to frame what’s happening: LLMs offer fluency and consistency at the expense of situated, idiosyncratic thinking. The metaphor fits. You get the same clean output everywhere, and what you lose is the local flavor, the unexpected angle, the reasoning style that doesn’t follow the dominant template.
And the political dimension is real. A few platforms, owned by multibillion-dollar corporations, control the algorithms and datasets driving this homogenization. Sourati et al. (2026) note the documented censorship in models like the Chinese Qwen, where refusal to answer politically sensitive questions illustrates how concentrated power can shape not just language but what counts as acceptable thought.
I also think we need to contextualize these findings carefully. This is an opinion paper, a synthesis of existing research, not an original empirical study. The authors are making an argument and marshaling evidence for it. The evidence is real, but the framing is deliberately provocative. Not every use of an LLM leads to cognitive homogenization, and the paper doesn’t always distinguish between casual use (asking ChatGPT to fix a typo) and deep reliance (using it to formulate arguments or make decisions). That matters for educators. The pedagogical question isn’t whether to use AI. It’s how to use it in ways that preserve the cognitive diversity Sourati et al. rightly value.
The technology will keep converging. The pedagogical challenge is to make sure the thinking doesn’t converge with it.
References
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing tasks. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
- Niloy, A. C., Akter, S., Sultana, N., Sultana, J., & Rahman, S. I. U. (2024). Is ChatGPT a menace for creative writing ability? An experiment. Journal of Computer Assisted Learning, 40(2), 919–930. https://doi.org/10.1111/jcal.12929
- Sourati, Z., Ziabari, A. S., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences. Advance online publication. https://doi.org/10.1016/j.tics.2026.01.003