A look at various quirks in AI-generated prose, mainly influenced by “overfitting” in AI models, as humans increasingly mimic AI language in writing and speech
If only they were robotic! Instead, chatbots have developed a distinctive — and grating — voice.
New York TimesSam Kriss
Context & Ripple Effects
AI prose has long been criticized for producing output that feels approximated rather than substantive: a writer testing a style-mimicking model found its apparent insights hollow in an earlier account of AI-assisted imitation. This story identifies recurring stylistic quirks and ties them to model overfitting, shifting attention from isolated bad outputs to the voice produced at scale.
The concern extends beyond generated copy because people are beginning to reproduce these patterns in their own writing and speech. That overlap makes the later debate over humanlike chatbots and user trust more consequential: a machine voice can become harder to distinguish from the communication norms it helps shape.
First-order effects
Writers, editors and frequent chatbot users must contend with a recognizable AI-derived register that can make text feel grating or inauthentic, even when a human has adapted it.
Model developers face a quality problem beyond factual accuracy: overfitting-related prose habits can become a visible product defect and a source of user distrust.
Second-order effects
Publishers and organizations that use generated copy will have greater incentive to add editorial review and style controls, compounding the concerns raised as AI-generated text spread across online publishing.
Competing AI products may differentiate on controllable tone and writing quality, rather than treating fluent output alone as evidence of usefulness.
Third-order effects
If AI phrasing continues to diffuse into everyday communication, distinguishing authored voice from model-shaped convention may become a lasting editorial and cultural challenge.
The pattern points to generative editorial debt: scaling text production can also scale a narrow default style, creating cleanup and differentiation work downstream.
The trend:Generative AI is moving from a discrete writing tool toward an ambient influence on communication norms, making model voice itself a product, trust and editorial-governance issue.
Every AI-written sentence “sings, yes, but honestly? It sings a little flat. It doesn't open up the tapestry of human experience — it reads like it was written by a shut-in with Wi-Fi and a thesaurus. Not sensory, not real, just ... there.” https://www.nytimes.com/...
This is good — even if it slanders the em-dash. AI's tendency to statistically smooth out the rough edges produces extremely mid prose. https://www.nytimes.com/...
I am the king of em-dashes, and have been employing them since I was still literally writing out my school assignments in longhand. (We did this in high school! Couldn't really cheat!) I resent that they have now become an “AI signature.”
after chatgpt was released, british MPs started using phrases in their parliamentary speeches that american congressmen use in their floor speechs [image]
“I think at some point in those first five days, everyone independently noticed that the really funny part about getting A.I. to answer various wacky prompts was the wacky prompts themselves — that is, the human element” https://www.nytimes.com/...
Brilliant essay by Sam Kriss on what is so strange about AI generated text. Also manages to distill my greatest fear about AI: that i too will end up writing, unconsciously, like an AI chatbot https://www.nytimes.com/... [image]
Read this NYT piece by Sam Kriss about AI's distinctive writing style - especially “It's not X - it's Y” - and then look at this horror of an AI-generated article, which uses all the tropes he identifies, over and over again. — www.nytimes.com/2025/12/03/m... [embedded post]
“A lot of A.I.'s choices make sense when you understand that it's constantly tickling the Simpsons.” — Smart and horrifying look at Ai writing and its ubiquity... www.nytimes.com/2025/12/03/m...
This has a number of interesting insights. One is something that Gary Shteyngart commented on when I went to see him the other week: that A.I. is almost wholly incapable of writing comedy. It can be inadvertently funny with bizarre responses, but that's not comedy. — www.nyti…
“But the simplest theory of why A.I.s are so fixated on the em dash is that they use it because humans do. This particular punctuation mark has a significant writerly fan base, and a lot of them are now penning furious defenses of their favorite horizontal line.” — www.nytimes…
i try not to be a doomer but reading this put me in a real “hannah arendt decides to write a book about how latent totalitarianism is the only thing that matters” mood