I won't read LLM authored fiction

The author describes how reading fiction changes his own writing by exposing him to the statistical profile of another writer's word choices. This exposure nudges his personal distribution of vocabulary and phrasing in a direction that feels fresher and more varied. He notes that this effect is especially strong with fiction because it tends to deviate from ordinary, utilitarian language. The mechanism is presented as a form of stylistic osmosis that improves his creative output.

He argues that large language models generate text by sampling from a learned distribution that captures typical usage patterns. Consequently, LLM-authored fiction exhibits a statistical profile that is close to the linguistic norm rather than to any individual writer's idiosyncrasies. Because the model's output is deliberately average, it lacks the distinctive variations found in human-crafted fiction. This average profile is what the author finds undesirable when seeking the stylistic boost he gets from reading.

Connecting the mechanism to his personal experience, the author says his own writing becomes easier and more fresh when he regularly reads human-written fiction. He feels that reading LLM-generated text would push his style toward the statistically normal, which is the opposite of what he wants from fiction. This aversion leads him to refuse any novel where most of the words come from an LLM, even if a human outlined the story at a high level. His stance is rooted in the desire to preserve the unique influence of human word-choice patterns.

For engineers, the observation raises a practical consideration: if they frequently consume LLM-produced documentation, code comments, or technical articles, similar statistical effects could influence the clarity and originality of their own writing. While the benefit of varied language may be less critical in strictly technical contexts, the risk of drifting toward bland, formulaic phrasing remains relevant. The piece encourages a mindful balance between using AI for efficiency and seeking human-authored material to maintain expressive richness.

The argument rests on the author's introspective account and lacks empirical validation or broader study. It may not generalize to all readers, different genres of fiction, or other forms of LLM output such as code or technical writing. Without data showing measurable shifts in word-choice distributions, the claim remains a personal perspective rather than a proven phenomenon. Engineers should weigh this anecdotal insight against their own experiences and any available research on language model exposure.