How to Write with an LLM: Methods and Boundaries
Use LLMs as writing collaborators, not ghostwriters — own your ideas, draft yourself, and always fact-check.
This article explores how to use large language models effectively in writing, arguing that LLMs should serve as collaborators, not ghostwriters. It breaks writing into stages, showing where models add clear value (ideation and review) and where humans must stay in control (drafting). It warns of two key risks: stylistic convergence toward a bland average, and confident-sounding hallucinations that demand independent fact-checking. A five-step workflow is offered, from forming your own ideas first to verifying every factual claim before publishing.
From Tool to Collaborator: Rethinking LLM-Assisted Writing
Large language models (LLMs) are changing how many people write. An article titled How to Write with an LLM, which sparked discussion on Hacker News, attempts to answer an increasingly practical question: when you have a tool capable of generating fluent prose right at your fingertips, how do you use it well — without letting it do your thinking for you?
The key distinction here is this: an LLM should be a collaborator in the writing process, not a ghostwriter. Handing an entire article over to a model typically yields something structurally tidy, stylistically safe, and utterly devoid of genuine perspective or personal voice — a "model answer" in the most literal sense. The truly valuable use is to involve the model in various stages of thinking, organizing, and refining, while keeping the human in control of direction and judgment.
A Sensible Division of Labor
Writing can be broken down into several stages: ideation, organization, drafting, revision, and polishing. LLMs vary considerably in how useful they are at each stage.
Where They Excel
During ideation, LLMs make excellent "thought starters." When you're staring at a blank page with no idea where to begin, you can ask the model to list multiple angles on a topic, raise counterarguments, or help break a vague idea into a few concrete sub-questions worth exploring. This kind of divergent assistance doesn't replace your judgment — it expands your thinking.
LLMs are also genuinely useful during revision. They can flag logical gaps, catch redundant or wordy sentences, and point out where an argument lacks support. Treating the model as a tireless first-pass reader is far more valuable than asking it to rewrite things outright.
Where You Should Be Careful
Drafting, ironically, is the stage you should least outsource to a model. How you organize your words, what you choose to emphasize, what you decide to leave out — these choices are themselves acts of thinking. If you skip this step and simply accept the model's output, you've effectively surrendered ownership of the content, and the ideas it expresses no longer truly belong to you.
Keeping Your Own Voice
A recurring concern in the Hacker News discussion was this: over-reliance on LLMs causes writing to converge. When thousands of people use the same models with similar prompts, the resulting text gravitates toward a kind of average — smooth, inoffensive, and stripped of any real edge.
The antidote is to treat LLM output as raw material, not a finished product. Ask it for a draft or some fragments, then rewrite them in your own words, add specific examples and personal experience, and cut anything that sounds templated. The most compelling writing tends to contain details and judgments that only you could have produced — and that's precisely what a model cannot do for you.
This "averaging" tendency has roots in how these models are built. Mainstream LLMs are trained on massive amounts of internet text and then fine-tuned via reinforcement learning from human feedback (RLHF), a process in which human raters tend to favor responses that are fluent, measured, and non-controversial. Over time, models develop a "lowest common denominator" style: clear paragraph structure, careful hedging, no rough edges. This is why articles produced by different users with similar prompts often end up sounding structurally and stylistically identical. Understanding this mechanism helps explain why asking a model to "preserve the original tone and wording" during editing is often not enough. The more fundamental fix is to run its output through your own voice one more time after the fact.
Fact-Checking Is Non-Negotiable
No matter how credible an LLM's prose sounds, it can and will "hallucinate" with complete confidence. Generative models are fundamentally predicting the most likely next token — not verifying facts. Any content involving specific data, citations, names, or events must be independently verified by a human.
Making this a fixed part of your workflow matters: treat any "factual" claim from a model as an unverified hypothesis until you've confirmed it yourself. This isn't just about being responsible to your readers — it also protects your own or your organization's credibility from damage caused by misinformation.
This phenomenon is technically known as "hallucination." LLMs learn statistical patterns in language during training; they do not store or reason about real-world facts. When a model encounters something it's "uncertain" about, it doesn't stop and say "I don't know" — instead, it generates a statistically plausible-sounding answer based on context. This can mean fabricating citations to papers that don't exist, providing incorrect data or dates, or even inventing quotes attributed to real people. Research suggests that a model's level of expressed confidence bears no meaningful correlation to its accuracy: it answers "Apollo 11 landed on the moon in 1969" with the same assured tone it uses to state a completely wrong historical detail. This is precisely why independent verification cannot be skipped, even when the model's references appear authoritative.
A Practical Workflow
Drawing on all of the above, here is a pragmatic LLM-assisted writing workflow:
- Think first, on your own: Before opening a chat window, jot down the core point you want to make in a few sentences. This ensures the article's "backbone" comes from you.
- Use the model to expand: Let the LLM suggest additional angles, raise challenges, and help organize your material — but don't let it determine your position.
- Draft the key parts yourself: Write the core paragraphs by hand, especially those where your argument is sharpest.
- Use the model to review: Hand your draft to the model and ask "where is the logic unclear?" or "what could be more concise?" — not "please rewrite this."
- Fact-check and finalize: Independently verify all factual claims, cut any templated phrasing, and make sure the whole piece sounds like you.
Closing Thoughts
LLMs won't make writing meaningless — but they do shift where "the human" stands in the process. The more powerful the tool, the more precious human judgment and personal perspective become. The true measure of using an LLM well isn't how quickly you can produce an article, but whether the final output still clearly carries your thinking. Keep the model in the role of collaborator, hold on to ideation and judgment as your core responsibilities, and writing can actually become both more efficient and higher in quality.
Related articles
NVIDIA cuML Accelerates Spectral Clustering: 100x+ Speedup Over Scikit-Learn Benchmarked
NVIDIA cuML lets Scikit-Learn spectral clustering run on GPU without code changes, delivering 200x+ speedups on large datasets. Learn how it works and how to use it.
Google DeepMind Launches New Institute to Bring the AGI Debate into the Open
Google DeepMind has launched a new institute to bring AGI debate into the public sphere. We analyze what this signals about the shift from technical competition to AI governance.
How Cooley Is Reinventing IPO Legal Workflows with ChatGPT: A Look at the GO Public Tool
Cooley built GO Public on ChatGPT Work to accelerate IPO legal workflows, helping lawyers catch issues earlier and focus judgment where it matters most.