AI learns language from skewed sources. That could change how we speak and think (www.theguardian.com)

Large language models (LLMs) have developed from training primarily on written texts such as books, social media, and scripted speech from movies and television, which creates a skewed representation of human language. This limitation poses significant risks because as AI-generated text becomes more prevalent, humans may unconsciously adopt the linguistic patterns of these models, leading to a fundamental shift in how we communicate and think. Research indicates that over-reliance on AI can simplify our language, erode expression, and foster a communication style that is curt and transactional, mirroring the brevity typical of digital interactions. The implications are profound: as AI increasingly reflects and reinforces a narrow slice of human expression, it risks promoting biases and distorting our understanding of the world. For instance, the confident tone of AI-generated text may contribute to overconfidence in unexamined thoughts while nurturing negative social behaviors learned from online interactions. Moreover, as LLMs create a feedback loop by training on their own outputs, they could distort the very fabric of conversation, reducing emotional richness and welcoming confirmation bias. To cultivate a more complete understanding and representation of human communication, there is an urgent need to explore ways to incorporate informal, natural speech into AI training datasets.