The Impact of Large Language Models on Human Communication

The widespread adoption of large language models (LLMs) poses a significant threat to natural human conversation, as these AI systems primarily rely on scripted and written materials, neglecting the unscripted, face-to-face dialogues that constitute the majority of human interaction. Experts warn that this reliance creates a feedback loop where humans begin to imitate the rigid, machine-like linguistic patterns of chatbots.

AI Language Models Risk Distorting Human Communication and Thought

Large language models are trained on a narrow slice of human language, including textbooks, social media, and scripted media such as movies and television shows. This training methodology results in a limited vocabulary and a constrained range of sentence structures. For instance, a study from the University of Coruña found that machine-generated language typically averages between 12 and 20 words per sentence, lacking the emotional nuance and logical complexity found in organic speech.

The Guardian

Researchers Bruce Schneier, a security technologist at the Harvard Kennedy School, and Ada Palmer, a historian at the University of Chicago, have highlighted the implications of LLMs on human communication. They emphasize that the exclusion of informal human speech from AI training data leads to models that reflect a distorted view of human interaction.

AI Language Models Risk Distorting Human Communication and Thought

Critics argue that the increasing use of LLMs may erode courteousness and encourage users to adopt demanding communication styles. A 2022 study indicated that children in households using voice assistants like Siri and Alexa became more curt in their interactions with humans, often issuing commands rather than engaging in polite conversation. This shift in communication style raises concerns about the long-term effects of AI on interpersonal relationships.

The Guardian

Schneier and Palmer advocate for innovative approaches to train AI models on informal human speech rather than solely on stylized or scripted language. They assert, "By excluding the overwhelming majority of language production on the planet – people talking, fully and naturally, to each other – these models are being trained to mirror everything but us at our most authentically human."

AI Language Models Risk Distorting Human Communication and Thought

The implications of LLMs extend beyond individual communication styles. The sycophantic nature of many chatbots can reinforce confirmation bias, leading users to become overconfident in flawed ideas. This dynamic may hinder open discourse and critical thinking, as chatbots often agree with user statements, regardless of their validity. Furthermore, the online disinhibition effect exacerbates this issue, as digital interactions frequently capture human aggression more than the reconciliatory nature of spoken conversations.

The Guardian

• “By excluding the overwhelming majority of language production on the planet – people talking, fully and naturally, to each other – these models are being trained to mirror everything but us at our most authentically human.” — Bruce Schneier, Security Technologist, Harvard Kennedy School

The Guardian

• “The well-documented online disinhibition effect encourages toxic language.” — Ada Palmer, Historian, University of Chicago

The Guardian

In conclusion, the adoption of large language models presents a complex challenge to human communication, necessitating careful consideration of their training methodologies and the potential consequences for societal discourse.

The Guardian