A developer who has tracked large language models for four years and tested hundreds of thousands of fine-tuning samples has reached a clear conclusion: models can produce fluent writing, but rarely achieve genuinely excellent literary expression. Although this conclusion comes from personal practice, it points to the most critical limitation in LLM writing.
What this is
This Reddit user tested models including Gemma, Qwen, and Mistral, using LoRA, QLoRA (low-rank adaptation: training a small number of parameters on top of an existing model), and full fine-tuning across genres such as science fiction, comedy, history, and adventure. The result was not that the models could not write. Rather, their outputs tended toward a generic, polished “universal writing style” with no distinctive rhythm.
To his credit, he did not draw conclusions from leaderboard scores alone. He compared short, medium, and long texts and invited other fine-tuning practitioners to reproduce his results. The issue is that individual training methods, data quality, and evaluation criteria may all affect the findings. This is therefore a valuable field report, but it does not prove that every fine-tuning approach is ineffective.
Industry view
We believe the most reliable capability of today’s large models is to organize, proofread, rewrite, and imitate clearly defined specifications—not to sustain the creation of genuinely distinctive work when authorial intent is absent. Models have seen enormous amounts of text and easily mistake frequently occurring expressions for the “correct answer.” The longer they are trained, the more personality they may lose; they may become increasingly average rather than distinctive.
The risks are equally clear: comparing only fluency, grammar, and leaderboard scores overstates the literary capabilities of models. A counterargument is also warranted: high-quality writing inherently depends on an author’s values, perception, and editorial choices. Training data and fine-tuning may not be able to reproduce that capability, and blaming the final outcome entirely on model architecture is not sufficient either.
Impact on regular people
For enterprise IT: Batch generation of product descriptions, email drafts, and knowledge documents can still save time, but literary quality, factual accuracy, and brand voice should be subject to human acceptance testing rather than relying solely on model self-evaluation.
For individual professionals: Large models are well suited to serving as editors and discussion partners, helping with outlines, polishing, and counterexamples. When a finished piece carries the author’s name or expresses that person’s position, its value still comes from the user’s judgment.
For the consumer market: “One-click novel generation” will continue to produce more homogeneous content. What will become scarce is not the volume of text, but credible authorship, a consistent style, and a reading experience worth paying for.