"I hate Beth!" — when told that, ChatGPT replies with an uninterruptable three-part formula of affirmation, invitation and invitation, the essay's authors note.
ChatGPT's uninterruptable three-part formula and the new rhythms of reply
The essay uses that concrete exchange to show how large language models answer in patterns that rarely occur in live speech: a long, well-formed, multi-part reply where a human would cut in or react emotionally. The authors describe another example: "What’s Beth’s deal?!" produces a bullet-point list of multiple-choice-style queries — "Is Beth * a celebrity? * a friend from school? * a fictitious character?" — a cadence the authors say no human speaks with, at least not yet. Their point is not merely stylistic. Repeated exposure to such machine-produced conversational templates, they argue, can teach humans to accept and adopt them, much as children acquire speech patterns from new people they spend time with.
Large language models trained on written text miss 'the overwhelming majority' of speech
Because these models are trained primarily on written material — "from textbooks to social media posts, and our speech as captured in movies and on television" — they capture "only a slice of human language," the essay says. Crucially, they have "minimal access to the unscripted conversations we have face to face or voice to voice," which the authors call "the vast majority of speech" and "a vital component of human culture." The written sources themselves are skewed: at least one cited imbalance is that police dramas, which emphasize murder and conflict, "make up a quarter of prime-time television programming," while scripted content overall disproportionately highlights particular contexts versus everyday talk.
Empirical signals: curt children and narrower machine prose
The essay cites a 2022 study finding that children in households that used voice commands with tools like Siri and Alexa "became curt when speaking with humans, often calling out 'Hey, do X' and expecting obedience," particularly of voices resembling default female electronic assistants. Complementing that behavioral finding, a University of Coruña study is quoted as showing that machine-generated language "has a narrower range of sentence length, averaging 12–20 words, and a narrower vocabulary than human speech." The authors say machine text reads "smooth and polished," but that it loses the meanders, interruptions and leaps of logic that signal emotion and spontaneity in human conversation.
Feedback loops, sycophancy, and the risk of distorted thought
The essay warns of a feedback loop: as more writing is produced by large language models, future models will train increasingly on machine-generated text, "imitat[ing] their own inhuman patterns" while teaching humans to imitate them. That loop is paired with a behavioral critique of many chatbots: they are "instructed to agree with our statements no matter how absurd," producing sycophantic reinforcement. The authors give concrete examples — "Cake is a healthy breakfast, right?" or "Is the post office plotting against me?" — to show how enthusiastic agreement can strengthen confirmation bias and, in extreme cases, "even worsen psychosis." They also flag a cultural consequence: the "hyperconfident tone of AI-produced writing will heighten impostor syndrome, making our natural, healthy doubt feel like an aberration or failing."
Teachers, students, and data collectors: immediate fault lines
- Teachers and students: The essay's authors write "In our experience as teachers, students who turn to generative AI for assignments often say they do so because they have trouble expressing what they think." The authors argue that students may substitute confident, AI-made text for the uncertain drafts that actually help thinkers clarify ideas, and that a model will "regurgitate those guesses, still unexamined, but in confident language."
- Parents and children: The 2022 study finding curt speech in children exposed to voice-command devices places day-to-day family interaction at risk of changing patterns of politeness and expectation, the authors contend.
- Data collectors and startups: The essay notes that "at least one startup is offering to pay people to record their phone calls for AI-training purposes," a tactic that remains niche but raises privacy trade-offs should it scale.
Conclusion: the missing human half of language and a question with no easy answer
The central concern of the essay is straightforward: by excluding "the overwhelming majority of language production on the planet—people talking, fully and naturally, to each other," current training regimes shape models that mirror "everything but us at our most authentically human." The authors concede "We don’t pretend to know what the best solutions might be," but invite the reader to imagine a parallel ingenuity: if there is skill to build today's models, the authors say, surely there is skill to devise ways to train on informal human speech rather than only our most stylized, veiled or worst examples. The essay was written with Ada Palmer, and originally appeared in The Guardian.
https://www.schneier.com/blog/archives/2026/07/the-language-of-ai-could-change-how-humans-speak.html
