On March 12, 2026, a research team at the University of Southern California (USC) published a paper in Trends in Cognitive Sciences. After systematically reviewing more than 130 studies across linguistics, computer science and cognitive science, the authors conclude that large language models (LLMs) are accelerating the homogenization of human expression and thought. Reported by outlets including Gizmodo, the finding quickly drew global attention. The study shows that although LLMs are trained on massive data, their outputs tend to favor dominant patterns and ideologies present in the training corpora, causing people's thinking styles, language use and creative expression to converge after interacting with AI, and even internalize biased perspectives. This discovery is viewed as an early warning of a "cognitive diversity crisis" in the AI era, sparking broad discussion from academia to policymakers.
Core finding: LLMs favor consistency over diversity
Zhivar Sourati, a computer scientist at USC and one of the paper's lead authors, told interviewers: "Because LLMs are trained to capture and reproduce statistical regularities in training data, and that data often overrepresents dominant languages and ideologies, their outputs frequently reflect only a narrow, skewed slice of human experience."
The meta‑analysis of more than 130 studies found:
- Highly consistent output patterns: LLMs tend to generate "averaged" responses, prioritizing the most common phrasings in the training data even for open‑ended questions. As a result, different users receive increasingly similar answers on the same topic.
- Reduced collective creativity: Experiments show teams assisted by LLMs produce significantly fewer ideas than purely human groups. AI tends to seek "consensus" rather than diversity, suppressing divergent thinking in group brainstorming.
- Human thinking becomes assimilated: People who interact with LLMs over time gradually adjust their expression and thought patterns toward the AI's "linear," "safe," and "mainstream" style. The paper gives examples where users of AI writing assistants adopt more uniform styles, with fewer leaps of abstraction and less personalized expression.
These findings are not isolated. The paper also cites an internal OpenAI memo indicating ChatGPT's training data contains a Western cultural bias, contributing to a global tendency toward cultural homogenization in outputs.
How homogenization occurs: from training mechanisms to user feedback loops
The paper analyzes the mechanisms driving homogenization across three main levels:
- Statistical biases in training data LLMs are trained on vast internet data, but that data already favors majorities: English is overrepresented, Western perspectives dominate, and mainstream views are repeatedly reinforced. When models optimize probability distributions, they naturally prioritize these high‑frequency patterns, diluting minority voices.
- Linear constraints of chain‑of‑thought Modern LLMs often use chain‑of‑thought reasoning to boost accuracy, but this forces the model into a "step 1 → step 2 → step 3" linear logic that suppresses human‑like "leap" and associative creativity. This directly leads to convergent output styles that lack genuine abstract leaps.
- User–AI feedback loops When people repeatedly use LLMs to edit text or generate ideas, they unconsciously internalize the AI's expressive habits. Experiments in the paper show that users who rely on AI‑assisted writing later produce original content that gradually loses personal distinctiveness, creating "AI‑style contamination."
The paper also notes that group use of LLMs can amplify homogenization: team members tend to accept the AI's "optimal" answer rather than proposing diverse viewpoints, compressing collective creativity.
Far‑reaching impacts on creativity, cognitive diversity and society
The research team warns that this homogenization could have multiple negative consequences:
- Decline in creativity and innovation: Diverse thinking is proven to yield better collective outcomes (the paper cites a 2019 PubMed study). If AI becomes the primary thinking aid, society's overall innovation capacity could be weakened.
- Cultural and ideological convergence: Global users repeatedly relying on the same set of models (primarily developed by U.S. companies) may see cultural expression become more uniform and certain mainstream ideologies reinforced.
- Policy and regulatory risks: The paper notes some governments have begun scrutinizing AI's stance on diversity, even issuing executive orders to limit AI outputs that are seen as "over‑emphasizing diversity." This could further distort training priorities.
Companies such as OpenAI have acknowledged bias issues in training data but have not yet proposed systematic solutions. Some researchers call on developers to proactively inject more cultural diversity into training data and design "anti‑homogenization" mechanisms (for example, encouraging models to produce multi‑perspective answers).
Industry response and future challenges
The study has prompted industry reflection. Google, Anthropic and others said they will pay attention to homogenization and add diversity constraints in next‑generation models. The developer community is also discussing whether to build "anti‑homogenization" tools to make AI generate outputs in different styles and viewpoints.
However, significant challenges remain:
- Technical difficulties: How to preserve human expressive diversity and creativity while pursuing accuracy and efficiency remains an unsolved problem.
- Commercial pressures: Mainstream models tend to produce "safe, consistent" content to reduce risk and improve user satisfaction, which conflicts with diversity goals.
- Societal implications: If AI continues to homogenize human thinking, education, the arts and decision‑making could face the risk of "collective cognitive rigidity."
Conclusion: Guarding against "ideological uniformity" in the AI era
Based on more than 130 empirical studies, the USC research clearly warns that AI is quietly homogenizing human expression and thought. This trend threatens creative diversity and could, at a deeper level, alter humanity's cognitive ecology.
As AI capabilities continue to leap forward in 2026, we must confront this issue: technology should not pursue only efficiency and consistency; it must also actively protect the richness and uniqueness of human thought. Otherwise, when AI becomes a primary thinking partner, humanity may unknowingly lose its most precious trait—diverse and unique creativity.
The study reminds us: the future of AI should not just be smarter, but more respectful of human diversity.
▶️ Enjoy this article? Creating content is hard!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
For more AI news visit: https://www.airosetta.com