How to Teach a Language Model to Write Indistinguishably From a Human Newsroom

The prompt "write like a journalist" doesn't work. That is the first thing anyone learns when they try to use language models to generate content for a specific outlet. The text comes out smooth, grammatical and sometimes even engaging, but it doesn't sound like the target publication. It sounds like ChatGPT pretending to be a journalist. A team of developers set out to solve exactly this problem, and their detailed technical write-up has appeared on Habr.

The post's author, Lena, for whom this is a first publication on the platform, describes the task with disarming honesty. The goal was not to make the model write "well." The goal was to make the text indistinguishable from something written by a specific newsroom: a specific city portal, a specific Telegram channel, a specific niche publication. The gap between those two formulations is a chasm that swallows most attempts at automating content.

Why is this hard in the first place? A publication's style is not a set of rules you can write into a system prompt. It is hundreds of implicit patterns: sentence length, how often colloquial phrases appear, a preference for certain syntactic constructions, characteristic ways of opening and closing paragraphs, even typical "imperfections," such as a newsroom's habit of overusing dashes or putting a period after every list item. By default, language models average all of this into a "generic good text" that belongs to no one.

The naive approach, a detailed prompt describing the style, hits a ceiling almost immediately. You can write: "use short sentences, a conversational tone, start with a provocative question." The model will dutifully follow the instructions, but the result is a caricature, not an imitation. It is like asking an actor to play a "sad person": you get the stereotype of sadness, not a particular sad individual. A style prompt describes a genre, not a voice.

The next logical step is few-shot examples, where the model gets several sample texts from the target publication directly in the request context. This works noticeably better but creates new problems. The context window is finite, and the more examples you load, the less room remains for the task itself. The model also starts copying specific phrases and facts from the examples instead of abstracting the style. It memorizes the surface, not the structure.

The approach that actually works, which the team arrives at, sits at the intersection of several methods. Fine-tuning on a corpus of the publication's texts lets the model absorb stylistic patterns in its weights rather than in its context. But pitfalls remain here too: you need enough data, you need careful filtering, and, most interestingly, you need metrics that measure stylistic similarity rather than just text quality. Standard benchmarks such as perplexity or BLEU score are useless here. The team developed its own metrics that analyze the distribution of sentence lengths, lexical diversity, the frequency of stylistic markers and other parameters that together form a publication's "fingerprint."

This case is interesting beyond its technical side. It raises a fundamental question about the future of media: if a model can be taught to imitate a newsroom's style indistinguishably, what does that mean for the very idea of an editorial voice? On one hand, it is a powerful scaling tool: a small newsroom can produce more content while keeping its stylistic integrity. On the other, it blurs the line between authorship and imitation. If a reader can't tell a model's text from a journalist's, who is the author?

There is a practical side too. The content market is already flooded with generic AI text that all sounds alike. Publications that can preserve a distinctive voice, even with the help of finely tuned models, gain a competitive advantage. The paradox is that a technology that threatens to make content faceless may become a tool for preserving its individuality.

The Habr post is, in essence, open documentation of an approach that many media companies are developing behind closed doors, and that openness is what makes it valuable. Style imitation will only get harder as publications start demanding not just competence from AI tools but character. Those who learn to solve this systematically, rather than through endless prompt rewriting, will set the standard for AI content quality for the next few years.