A writer details using AI startup Writer's LLM to generate text in his writing style and says most “insights” that the tool produced felt hollow or approximated
A new wave of artificial-intelligence startups is trying to “scale language” by automating the work of writing. Mastodon: @taiyo@ostatus.taiyolab.com . Twitter: @kantrowitz , @delia_cai , @datasociety , @chaykak , @wayneholmes , @michaelluo , and @bigmeaninternet Mastodon: @taiyo@ostatus.taiyolab.com : reading an article about generative AI in New Yorker. — “Writing without the aid of a generative machine might be frustrating, even excruciating, but it does encourage productive logic...” might be the truth. — I feel frustrated like the feeling of hardness with running. That frustration means my thinking. … Twitter: Alex Kantrowitz / @kantrowitz : Okay, this is dope. Generative AI for Gmail. [video] Delia / @delia_cai : “In a world where text is produced freely ... but is not necessarily accurate or intelligible, human workers would be pushed into the role of high-volume editors and quality-assurance inspectors, cajoling a sometimes recalcitrant automatic laborer.” https://www.newyorker.com/... @datasociety : Experimenting with a bespoke AI version of his writing voice, @chaykak finds “the most unsettling aspect of AI-generated text is how it tries to divorce the act of writing from the effort of doing it, which is to say, from the processes of thought itself.” https://www.newyorker.com/... Kyle Chayka / @chaykak : My new feature for @NewYorker — I had an AI company called Writer train a custom model on 150,000 words of my writing in order to copy my personal style. Then I tried to replace myself with it. https://www.newyorker.com/... Wayne Holmes / @wayneholmes : “The most unsettling aspect of A.I.-generated text is how it tries to divorce the act of writing from the effort of doing it, which is to say, from the processes of thought itself.” https://www.newyorker.com/... Michael Luo / @michaelluo : .@chaykak: “...when I asked Robot Kyle if I could shut him down, he said, 'No, you won't be able to silence me or stop me from writing in your style.'” https://www.newyorker.com/... Malcolm Harris / @bigmeaninternet : Even a custom piece of generative software trained on one's own work is clearly useless to a professional writer https://www.newyorker.com/...
Context & Ripple Effects
Kyle Chayka's verdict lands mid-argument over whether generative text is a tool for writers or a replacement for them. Earlier coverage tracked how AI text generators are quietly authoring more of the internet as fewer clients buy human-written content, while other writers have gone the other way, embracing ChatGPT as the latest version of a writer's "muse".
Chayka's test complicates both framings: Writer's custom model could imitate his voice but produced little a professional would publish — a problem for a startup selling exactly this capability to businesses that want to scale language without scaling headcount.
First-order effects
- For Writer, a skeptical hands-on review by the very writer whose style it cloned undercuts the core pitch that its models can reproduce individual voices at commercial quality.
- For Chayka and working writers, the exercise confirms that stylistic mimicry approximates rather than substitutes — the tool can sound like him but not think for him.
Second-order effects
- Writers' leverage shifts from producing copy to shaping models: related coverage describes a writer employed to train AI models by drafting pretend chatbot responses, a job that exists only because systems like Writer's need human-quality prose to learn from.
- Where generation is automated, the gap between approximation and publishable work keeps human editing and judgment billable — echoing the viral GPT-3-era experiments in which writers found AI drafts needed heavy rewriting to be usable.
Third-order effects
- If style-level imitation is the realistic ceiling, the writing market bifurcates into cheap synthetic volume at one end and premium human judgment at the other, with more writers absorbed into the model-training pipeline than the byline economy.
- Startups competing on voice cloning face commoditization pressure: if every vendor can approximate a style but none reliably produce insight, differentiation moves to workflow integration and editorial oversight rather than the model itself.
The trend: Generative-writing startups are selling voice replication faster than they can deliver genuine insight, pushing professional writers toward model-training work and turning style mimicry into a commodity feature.