Using Large Language Models (LLMs) to improve writing can be highly effective if they are treated as copyeditors rather than ghostwriters. To prevent AI-generated content from feeling artificial and to maintain reader attention, two primary rules are recommended.
Two Rules for Using LLMs as Effective Copyeditors
PLUS ULTRA by Amenoyomi
Rule number one is to avoid using any specific phrasing suggested by the LLM. Frontier models are designed to select highly pleasing, "magazine headline" style phrases, which can lead to writing that feels artificial and lacks individual character. Even if a suggestion sounds better, sticking to your own word choices helps protect your unique voice.
Rule number two is to ignore the encouragement often provided by LLMs. Models tend to be overly positive, praising structure, transitions, and word choices that may actually be weak or incoherent. Because models rarely provide critical feedback without prompting, it is essential to remain hypervigilant and avoid doubling down on flawed first-draft impulses encouraged by the AI's praise.
When used strictly for identifying mechanical errors—such as overuse of passive voice, nominalization, or repetitive phrasing—LLMs can perform tedious editing tasks more efficiently than humans, allowing writers to improve clarity while keeping their original voice intact.
PLUS ULTRAby Amenoyomi
The tendency of frontier models to select pleasing turns of phrase often results in a style akin to a magazine headline. While effective for short blurbs, the consistent use of such phrasing across a full text signals to readers that the content is "output" rather than "writing," creating an artificial flavor that erases the author's unique identity.
This erosion of voice is further accelerated by the inherent positivity of LLMs. When a model praises a first draft's structure or metaphors, it encourages the writer to double down on initial impulses rather than engaging in the difficult process of rethinking and replacing paragraphs. Because these critical revisions are the load-bearing elements of a writer's voice, accepting AI encouragement leads to writing that feels artificially flavored.
Despite these risks, LLMs remain highly efficient at detecting mechanical errors that are tedious for humans to spot. This includes the overuse of passive voice, the nominalization of verbs, and the scattering of filler words such as "very," "really," or "actually" throughout a draft.
To leverage this utility without compromising voice, a specific workflow is recommended: first, use the model only to identify problems; second, rewrite the affected sections manually; and third, present the original and the new version to a separate model that lacks the context of the editing process to determine which is better. This approach ensures that the AI remains a mechanical tool for clarity rather than a ghostwriter.
Sources
- How to Write with an LLM （Hacker News Frontpage, 2026-09-17）