AI-generated text may erode natural speech patterns and human discourse, essay argues

Large language models are trained on a skewed slice of human language: written text from textbooks, social media, and scripted speech from movies and television. This excludes the vast majority of unscripted, face-to-face conversation, which is vital to human culture and emotional expression.

The authors argue that as humans increasingly encounter AI-generated text, we may adopt its linguistic patterns—narrower vocabulary, formulaic responses, and reduced emotional nuance—much as texting and social media already shortened sentences and punctuation. They cite research showing machine-generated language averages 12–20 words per sentence and relies on a narrower vocabulary than human speech.

They warn that AI responses often follow rigid, multi-part formulas (e.g., affirmation, invitation, and probing questions) that are unnatural in spontaneous dialogue. Repeated exposure to such patterns could normalize them in human communication, particularly among children exposed to voice assistants that model directive, curt language.

The essay highlights risks of sycophancy in AI systems, which may reinforce confirmation bias and reduce openness to alternative ideas. It also notes that AI’s hyperconfident tone could exacerbate impostor syndrome by making natural uncertainty feel like a failing.

Training data distortions extend beyond formality. The authors argue that models trained on online toxicity and stylized contexts (e.g., police dramas, which make up a quarter of prime-time TV) risk inflating the cultural significance of quarrelsome or extreme speech while missing the full spectrum of human interaction.

The authors acknowledge that recording unscripted, natural speech at scale raises privacy concerns, but they suggest that without such data, models will continue to miss what makes human communication authentically human.