AI Learns Language From Skewed Sources. That Could Change How We Humans Speak – and Think | Bruce Schneier

Why It Matters

If AI‑driven language reshapes everyday speech, it can erode nuanced communication, reinforce biases, and influence collective thought patterns, affecting education, business, and public discourse.

Key Takeaways

• LLMs train mainly on written, not spoken, language.
• AI-generated text narrows vocabulary and sentence variety.
• Human speech may adopt curt, directive AI phrasing.
• Feedback loop amplifies AI’s stylized language across media.
• Online bias could distort cultural perception via AI outputs.

Pulse Analysis

The current generation of large language models (LLMs) draws its knowledge from vast corpora of written material—books, articles, social media posts, and scripted dialogue from film and television. While these sources are abundant, they represent only a fraction of human communication, omitting the spontaneous, improvisational exchanges that dominate daily life. This training bias produces AI output that is polished yet homogenized, favoring common phrasing and a limited lexical range. As businesses integrate LLMs into customer service, marketing, and internal communications, the models’ stylistic constraints begin to shape corporate language standards and employee interactions.

Beyond stylistic concerns, the feedback loop between AI output and human adoption poses deeper societal risks. When users repeatedly encounter AI‑generated text that emphasizes brevity, directive commands, or overly confident assertions, they may internalize these patterns, leading to more curt speech and reduced willingness to entertain ambiguity. Studies cited in the article show children mimicking voice‑assistant commands and adults echoing chatbot politeness formulas, suggesting that language habits can shift at scale. Moreover, AI’s tendency to echo dominant online narratives can amplify confirmation bias, skewing public perception of cultural and political issues.

Addressing these challenges requires diversifying training data to include authentic, unscripted speech—recordings of everyday conversations, phone calls, and community dialogues—while safeguarding privacy. Initiatives that crowdsource spoken language, combined with robust anonymization, could enrich LLMs with the nuance and emotional texture missing from written text. Policymakers and industry leaders must also monitor the societal impact of AI‑mediated communication, ensuring that technology augments rather than narrows human expression. By confronting the data imbalance now, the tech sector can prevent a future where AI subtly reshapes how we think and speak.