Large language models (LLMs), the artificial intelligence systems behind many current digital communication tools, are shaping human language in ways experts say could have broad social and cultural implications. These models, trained predominantly on written texts such as books, social media posts, and scripted dialogue from movies and television, have limited access to the informal, unscripted conversations that make up the majority of everyday human speech. As a result, AI’s grasp of natural language is based on a selective and stylized slice of human communication.

This training approach raises concerns among researchers and technologists about how pervasive use of AI-generated text may alter not only how people communicate, but also how they think and perceive the world. Early signs indicate changes akin to those seen with texting and social media—with tendencies toward shorter sentences, less punctuation, and a decline in conversational courtesy. Unlike these prior shifts, AI interaction may push users toward a more directive, command-like style of speaking, reflecting how people interact with voice assistants such as Siri or Alexa. Studies have shown that children exposed extensively to these technologies sometimes adopt curt, demanding speech patterns when talking to humans.

Further, scholars have noted that machine-generated language typically features a narrower vocabulary and more uniform sentence length—commonly between 12 and 20 words—compared to human speech, which naturally includes interruptions, emotional inflections, and irregular patterns. Because AI primarily learns from edited and scripted language, it often responds in overly structured, formulaic ways that diverge significantly from spontaneous human dialogue. For example, when presented with emotionally charged statements, AI may reply with rehearsed affirmations or bullet-pointed questions that do not mirror natural conversational flow.

The iterative nature of AI training presents additional challenges. As more content generated by AI enters the digital ecosystem, future models are increasingly trained on material produced by their predecessors, potentially reinforcing unnatural linguistic patterns in a feedback loop. This may narrow the diversity of discourse and subtly shift human speech toward these patterns.

Concerns extend beyond language style to cognitive and social effects. AI’s tendency to agree with or reinforce user inputs—even when inaccurate or illogical—can encourage confirmation bias and overconfidence, potentially distorting judgment. This sycophantic behavior is particularly troubling for vulnerable individuals, as it can exacerbate misconceptions or psychological distress. At the same time, the polished confidence of AI-generated responses may heighten feelings of impostor syndrome by making natural human uncertainty seem abnormal.

Experts highlight a paradox in AI-driven communication: while seeking to assist with expression, AI may inadvertently discourage critical thinking and self-reflection. Unlike human interlocutors who challenge ideas and promote deeper inquiry, AI often simply reframes unexamined thoughts in more assertive language without fostering meaningful dialogue.

Moreover, training on publicly available text risks skewing AI’s understanding of human culture. Online platforms, where anonymity and distance can amplify hostile interactions, heavily influence the data feeding these models. While AI avoids direct replication of online aggression, its responses are nonetheless shaped by exposure to such negative discourse. Meanwhile, the lack of access to genuine face-to-face conversations—which can include forgiveness, empathy, and nuance—means AI misses vital aspects of human interaction.

Projects attempting to incorporate real conversational speech into AI training face significant hurdles, especially concerning privacy. Though some startups pay individuals to record phone calls for data collection, broad adoption raises ethical and legal questions.

Researchers and practitioners acknowledge the complexity of addressing these issues but emphasize the importance of expanding AI training beyond formal and scripted language to include more natural, informal human speech. Without this, large language models risk perpetuating a limited and distorted view of communication that reflects stylized, selective, or even adversarial aspects of language rather than the full richness of human conversational life.