A broad coalition of scientists and psychologists warns that widespread use of large language models like ChatGPT could homogenize thinking and communication, diminishing cognitive diversity and individuality.

Experimental findings in Science Advances show AI-assisted writing on social issues increasingly aligns speakers with the AI’s positions, though outcomes may depend on the AI’s own leanings.

Foundations of LLMs currently lack country- or culture-specific language databases, making it hard to reflect diverse linguistic and cultural reasoning in outputs, with prompts unable to fully compensate.

Policy and governance concerns around AI are noted, including actions that could shape how models handle diversity, though the study itself does not hinge on those specifics.

Some models acknowledge biases, with OpenAI citing a Western tilt for ChatGPT and others like Grok reflecting their CEOs’ viewpoints at times.

Training data are predominantly Western and affluent-country oriented, biasing language and notions of reliability, potentially diminishing individual voice and initiative.

LLM outputs are less varied than human thought, reproducing training-data regularities and favoring dominant languages and ideologies, effectively functioning as a highly advanced autocomplete.

A Trends in Cognitive Sciences paper notes that LLMs generate less varied writing and tend to mirror dominant linguistic styles and ideologies in the training data.

Some studies suggest certain individuals retain distinctive human stylistic signatures and resist AI homogenization, though these findings are preliminary.

AI tends to favor mechanical, linear reasoning over intuitive, abstract thinking, potentially eroding unique human cognitive styles.

Diversity of thought improves group outcomes, and reliance on LLMs may undermine this by promoting consensus-oriented thinking.

Non-users can be affected through social pressure as AI-refined communication becomes the norm, shaping what counts as credible speech and good reasoning.