AI learns language from skewed sources. That could change how we humans speak – and think | Bruce Schneier

Editor's take

Large language models, trained predominantly on vast internet text corpora, are learning distorted representations of human communication, a phenomenon that could subtly reshape how people interact and conceptualize ideas.

This matters because LLMs like OpenAI's GPT-4 and Google's Gemini are increasingly integrated into everyday tools, from search engines to writing assistants. Their skewed linguistic patterns, lacking the nuance and context of genuine human dialogue, risk normalizing a simplified or biased form of expression, potentially impacting everything from casual conversation to formal discourse.

Future developments to monitor include whether developers can create training datasets that better reflect the full spectrum of human language and interaction. The long-term impact hinges on whether users develop critical awareness of LLM-generated text or passively adopt its linguistic tendencies.