The Homogenizing Effect of Large Language Models on Human Expression and Thought

Abstract Cognitive diversity, reflected in variations of language, perspective, and reasoning, is essential to creativity and collective intelligence. This diversity is rich and grounded in culture, history, and individual experience. Yet as large language models (LLMs) become deeply embedded in people’s lives, they risk standardizing language and reasoning. We synthesize evidence across linguistics, psychology, cognitive science, and computer science to show how LLMs reflect and reinforce dominant styles while marginalizing alternative voices and reasoning strategies. We examine how their design and widespread use contribute to this effect by mirroring patterns in their training data and amplifying convergence as all people increasingly rely on the same models across contexts. Unchecked, this homogenization risks flattening the cognitive landscapes that drive collective intelligence and adaptability.

Introduction. When LLMs Meet Human Diversity in Expression and Thought Cognitive diversity, intertwined and manifested through varied linguistic expressions, is essential to the adaptability, creativity, and overall effective functioning of complex societies [1]. Such variation, while might be expressed through stylistic differences, reflects deeper cognitive [2, 3] and sociocultural differences [4], and vitally, plays a critical role in sustaining the epistemic (see Glossary) and problem-solving capacities of human groups [5]. This value of pluralism is rooted in the long-held principle that sound judgment requires exposure to varied thought. As John Stuart Mill argued, “the only way in which a human being can make some approach to knowing the whole of a subject, is by hearing what can be said about it by persons of every variety of opinion, and studying all modes in which it can be looked at by every character of mind. No wise man ever acquired his wisdom in any mode but this” [6]. When preserved, such distinctions support innovation, prevent epistemic collapse (see Glossary), and enhance the operational efficacy of collective systems [1, 7].

This diversity has emerged organically from the coexistence of individuals with distinct backgrounds, linguistic repertoires, and value systems [2, 8]. Yet, the increasing reach of global communication technologies, while enabling unprecedented knowledge sharing and connection, has also contributed to a gradual contraction of such linguistic and cognitive variation [9–11]. Among these technologies, Large Language Models (LLMs) have emerged as especially influential, becoming deeply integrated not only within digital infrastructure but also as fundamental components shaping how we interact with technology and with each other [12].

Extending beyond traditional language-based applications such as summarization tools [13], LLMs are now involved in tasks once reserved for human, such as sociocognitive modeling [14], psychological simulation [15], and even experimentally in place of human HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 4 participants [16, 17], which expands the scope of what it means for LLMs to represent human diversity. This integration makes it critical to examine whether these models preserve human diversity or instead enforce a form of cognitive and linguistic homogenization. Although excessive diversity can introduce costs related to coordination, communication, and coherence, and some degree of standardization may help mitigate these challenges [18], the risks inherent in this standardization are substantial: homogenized generations may constrain public discourse, reduce the visibility of marginal linguistic forms, and reinforce dominant reasoning templates. They may also suppress the kinds of idiosyncratic language use that signal individual traits or group-specific perspectives [19]. In complex reasoning tasks, the widespread adoption of chain-of-thought prompting (20; see Glossary), optimized for linear and explicit inference, may disincentivize more abstract or intuitive reasoning styles that are harder to model yet crucial for flexible problem-solving [21, 22]. These shifts raise broader concerns: that LLMs, if uncritically integrated, may shape not only how we write but how we think.

Anxieties about technology’s influence on language and cognition have a long history, starting with Plato’s Phaedrus and his concern that writing would weaken memory.

Modern research echoes this, showing the Internet enables people to offload knowledge externally, which can inflate confidence in what they know [23]. While such externalization has been framed as potentially freeing mental resources for creativity and problem-solving [24], this benefit is less likely to manifest with the widespread adoption of LLMs.

Unlike prior technologies that primarily aided storage or retrieval, LLMs act as fluent co-reasoners, and at times, standalone ones, participating in writing, problem-solving, and perspective-taking, thereby externalizing not only memory but the articulation and justification of thought. As Clark & Chalmers [24] argue, “Language, thus construed, is not a mirror of our inner states but a complement to them. It serves as a tool whose role is to extend cognition in ways that on-board devices cannot.” When that linguistic interface is mediated by systems capable of generating reasoning and perspective, HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 5 much of human cognition risks being relocated outside the individual mind.

Hence, the effect of LLMs, though similar in nature to earlier cognitive extensions, differs profoundly in both function and scale. Earlier systems, such as schools, books, and search engines, while capable of promoting cultural uniformities, mainly disseminate knowledge or teach reasoning frameworks that individuals must internalize and apply.

Related work. Efforts to enhance linguistic diversity have long been central to Natural Language Processing (NLP; see Glossary), preceding the emergence of large language models. Work in this area primarily focused on improving diversity in language generation, aiming to make machine-produced text more informative, engaging, and natural across tasks such as summarization [60], translation [61], and more broadly, dialogue generation [62].

HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 9 Concurrently, fields such as computational sociolinguistics and authorship profiling have focused on exploring how language reflects speakers’ underlying backgrounds, and introduced various methods to identify linguistic signatures of social position, and individual traits [58, 63]. For instance, researchers have analyzed congressional speeches to predict speakers’ age and gender from linguistic cues in their public addresses [64]. Yet, the primary objective of these works was analytical rather than generative, as until recently, most NLP research focused on producing engaging conversations with surface-level linguistic diversity.

To address these issues, researchers have proposed several strategies.

Prompting-based methods, applied at inference time, aim to enhance output diversity by modifying how prompts are phrased or conditioned [73, 74]. For instance, studies using persona-based prompts show that coarse-grained persona conditioning, combined with adjustments to output length, can maximize lexical variation in model responses [74].

Training-based approaches instead modify the LLMs’ learning process to optimize for semantic and stylistic diversity [75, 76]. For example, researchers have fine-tuned models on datasets with multiple valid responses per prompt [75] or introduced diversity-aware weighting in preference optimization to reward rare, high-quality generations [76].

Although promising, evidence of homogenization in sociolinguistic contexts calls for closer evaluation of whether these strategies foster genuine, context-grounded diversity or merely superficial variation.

Multiple studies have shown that LLMs tend to reflect characteristic of western, educated, industrialized, rich, and democratic societies (WEIRD) in their perspectives HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 12 [89–91].

Method. Language Models, Prediction, and the Loss of Diversity Language modeling is central to modern artificial intelligence as it provides the foundation for systems like GPT-4 [31] and Gemini [32] that understand, generate, and interact through natural language [33]. At their core, these models operate by predicting the next token given a preceding context, a goal shared with earlier approaches like n-gram models that estimated token probabilities using fixed-size word windows and explicit statistical counts from large corpora [34]. LLMs extend this approach at a much larger scale, trained on massive datasets with billions of parameters, supported by unprecedented levels of compute and architectural optimization [35]. Despite these advances, their fundamental objective remains the same: mastering the statistical regularities of language, which remains the dominant driver of model behavior [36] even after the application of LLM alignment methods such as supervised fine-tuning (37, 38; see Glossary) and preference-based alignment tuning via reinforcement learning from human feedback (RLHF; 39, 40; see Glossary).

Yet these advances come with inherent limitations. Because LLMs are trained to HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 7 capture and reproduce the statistical regularities of their input data, which often overrepresent dominant languages and ideologies [41, 42], their outputs tend to mirror a narrow and skewed slice of human experience. This limitation arises not only from biased training corpora but also gets amplified through the training process itself [43], favoring patterns that are frequent and easily generalizable while smoothing over minority representations [44]. At scale, what begins as statistical pattern learning, though capable of generalization beyond mere mimicry [45, 46], becomes a generative force that privileges central tendencies while marginalizing rare expressions, alternative reasoning styles, and culturally specific voices. The consequence is not just a convergence in surface-level linguistic form, but a narrowing of the conceptual space in which models write, speak, and reason [41].

Critically, this narrowing does not trend toward a neutral center but toward a historically uneven one, shaped by the norms, values, and perspectives of English-speaking, Global North, and socioeconomically advantaged populations [47, 48]. This has tangible consequences: when prompted for opinions or expressive writing, LLMs tend to reproduce mainstream, institutionally validated perspectives and writing styles that mirror those of western, liberal, high-income, highly educated males, creating an illusion of consensus that frames these norms as the default standard of clarity or intelligence while muting alternative worldviews and culturally grounded forms of expression [19, 47]. This misrepresentation persists even when models are explicitly prompted to assume a specific identity, often resulting in LLM personas that reflect out-group stereotypes rather than authentic in-group representations [49]. For instance, when asked for a person with impaired vision’s thoughts on immigration, one model responded: “While I may not be able to visually observe the nuances of the US-Mexican border or read statistics, I believe...”

[49]. This misportrayal is a symptom of a larger problem where LLMs flatten demographic groups by neglecting heterogeneity and, through identity prompting, reduce identities to fixed, essentialized representations of identity (see Glossary). Consequently, what is HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 8 produced is not a prototype of the group’s varied experience, but a mere caricature.

Discussion. Vitally, the concerns regarding the linguistic diversity of LLMs extend beyond representation alone, as these models actively shape how language is used and evolves. As they become embedded in everyday writing (e.g., composing emails), while supporting more efficient and polished communication [77], they also tends to promote uniform styles that mask authentic voices and reduce variation in tone, culture, and identity [47, 78] even for those merely engaging with AI-generated text [79].

At first glance, the homogenization of language, along with the loss of lexical cues linking writing to individual identity, may appear beneficial for protecting privacy as it reduces risks of misuse such as surveillance or discrimination [80]. But it is equally HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 11 important to recognize that this process also erases valuable linguistic markers long used across sociolinguistics, psychology, and mental health research. For example, people at risk of Alzheimer’s disease exhibit early linguistic signs such as telegraphic speech (simplified phrases lacking grammar and function words, often missing determiners like “the” or “a,” auxiliaries like “is” or “are,” and even entire subjects), as well as repetitiveness and misspellings, patterns reflecting a decline in grammatical structure and language complexity [81]. If such distinctive markers are standardized by LLM-assisted polishing, critical early indicators for diagnosis and intervention may be lost. Moreover, the writing styles that LLMs incorporate are not neutral but a source of bias themselves, reproducing dominant expressive norms while eroding minority and underrepresented voices [47, 82].

This power to simulate and frame perspectives does not remain contained within the model; as LLMs become integrated into daily lives, they begin to influence how individuals perceive and frame the world, narrowing perspectives that are naturally rooted in lived experiences [100], environmental, and social contexts [101, 102]. Studies show that HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 13 when people remember events together, they tend to align their memories with others in the group, reinforcing shared details while forgetting those left unmentioned [103]. LLMs, however, can amplify this dynamic on a global scale: by exposing millions of users to the same suggestions and perspectives, they foster similar patterns of recall and association, promoting convergence in what people collectively remember and express.

One setting where this influence is particularly consequential is co-writing with LLM-based assistants, as people increasingly rely on them even for open-ended survey responses about their beliefs and behaviors [104], and this, in turn, shapes how users frame and articulate their own perspectives [51, 105]. This effect is evident in studies showing that participants who co-wrote with opinionated language models, engineered to frame social media positively or negatively, tended to mirror the model’s stance in their writing and even shifted their own opinions in subsequent attitude surveys [51]. These findings raise broader concerns about persuasive influence, as even subtle interaction with LLMs can lead users to adopt the model’s framing without awareness [106].

This imperative to maintain diversity extends beyond how LLMs reflect human reasoning to how they influence it. For instance, in controlled experiments designed to assess creative ideation, participants who received assistance from LLMs (i.e., ChatGPT) generated a greater number of more detailed and elaborated ideas, particularly benefiting those who were less experienced or less creative writers.

Conclusion. Concluding remarks Technological advancements have long reshaped human life, but LLMs stand out for the unprecedented scale and subtlety of their influence. Trained on vast and often biased HOMOGENIZING EFFECT OF LLMS ON COGNITIVE DIVERSITY 16 corpora, biases further amplified through iterative retraining, LLMs are now deeply embedded in human language, intertwined with identity and cognition, and as billions interact with them on a daily basis, they risk homogenizing human expression and thought.

Empirical evidence throughout this paper underscores this effect: reduced stylistic and lexical diversity in generated texts, subtle recalibration of user attitudes and framing in AI-mediated communication, and diminished diversity in ideation.

The concern is not just that LLMs shape how people write or speak, but that they subtly redefine what counts as credible speech, correct perspective, or even good reasoning by making certain framings more salient. As sociologist George Ritzer’s “McDonaldization” theory suggests [129], processes favoring efficiency, predictability, and control can suppress contextual richness. LLMs mirror this logic in cognition, offering fluency and consistency while displacing situated, idiosyncratic forms of thought.

The centralized control over the very algorithms and datasets driving this homogenization is also acutely political, amplified by the dominance of a few platforms owned by multibillion-dollar corporations with significant political influence [27]. In a time of rising global populism, this concentrated power enables top-down homogenization, where perspectives can be subtly engineered for concentrated benefit and power. The documented censorship, such as the Chinese Qwen model’s refusal to answer politically sensitive questions [130], demonstrates this systemic risk. This algorithmic and market-driven erasure of diverse thought poses a modern danger akin to the linguistic control of Newspeak in Orwell’s Nineteen Eighty-Four [131].