Writing with AI

Writing with AI works best when the human remains the writer.

A model can find a better phrase, expose a weak transition, reorganize a draft, or show why a paragraph does not land. What it cannot know is which experience matters, what you truly believe, or which imperfect sentence sounds like you. Delegate those decisions and the work becomes something else.

This guide is about keeping that collaboration useful. Important! Writing with AI does not necessarily make writing faster. Sometimes it does. Sometimes the back-and-forth takes longer. Done well, it makes the text better whether or not it saves time.

The human writes, the model responds

Start with a blank page and a broad request such as “write an article about this,” and the model gets too much authority too early. It chooses the argument, the examples, and the shape of the piece before the writer has decided what to say.

That can feel productive. It can also quietly replace authorship with selection from mediocre text. The first draft from a blank prompt is usually rubbish, even from a strong model.

AI is not built to write great text. It is trained to produce the likely next word, which pulls it towards the statistical middle: polished, coherent, and forgettable. Think of it as the disciplined half of a writing pair. It handles grammar, tracks the facts you provide, and catches details. You have to be the mad one: curious, opinionated, and willing to try the idea that should not work. The model brings cold perfection. Useful, but far from enough. Without you, it is worthless.

This pattern has a name: reactive writing. Instead of developing an idea, the writer begins by judging suggestions and personalizing a direction the model introduced. The result still feels like theirs because they can edit it, even though the first suggestion has already shaped the argument. (Reactive Writers)

Repeated direct editing can also weaken a writer's sense of authorship. Asking the model for critique and feedback preserves more of it. (Authorship Drift)

A better division of work is simple:

• The human supplies experience, intent, taste, and judgment.
• The model supplies attention, alternatives, structure, and editorial distance.
• The text moves between them until it says what the human meant.

That back-and-forth is not friction to eliminate. It is where the writing happens.

A practical writing process

Start with material, not prose

Before asking for a draft, write down what you know:

• Who will read it?
• What should they understand by the end?
• What have you observed yourself?
• Which facts, examples, and distinctions must survive?
• What are you unsure about?
• What would sound false coming from you?

Rough notes keep the boundary visible. You can still see what came from the writer. A polished first draft tends to hide it.

Use bracketed editorial briefs

A simple way to direct the model is to mix headings, fragments, and instructions in square brackets:

## Crawlers and scrapers

[Explain both through a practical example.]
[The distinction matters, but do not turn this into a glossary.]
[Lead naturally into the section about search engines.]

Scrapers can copy prices, articles, or product data.

[This is too broad. Show that some automated access is useful.]

The brackets separate what the piece should do from what the piece should say. They also leave room for gaps without inviting the model to fill them with invented facts.

A useful brief might contain:

[Audience]
[Purpose]
[Facts and experiences to preserve]
[Points the reader needs in this section]
[Desired tone]
[What to avoid]
[Open questions]

Do not fill every field for every task. The format is a thinking aid, not a form.

Ask for criticism before a rewrite

“Improve this” often produces a smoother version of the same problem. Ask what is unclear, repetitive, misplaced, too abstract, or unsupported first. Then ask for replacement wording with every criticism. Suggestions are easier to judge in context.

The writer can then respond line by line:

> Suggested sentence

[Keep this idea, but the wording is too formal.]
[Remove the claim about cost.]
[Use my example instead.]

This is faster than requesting a complete rewrite every time. It also makes disagreement useful. Rejecting a suggestion teaches the editor something about the intended voice.

Work locally, then read globally

Most revisions should be small. Fix the paragraph that is wrong without disturbing what already works.

After several local edits, read the complete piece again. Good paragraphs can still make a poor article. Check whether:

• the opening creates the right question;
• each section prepares the next one;
• ideas appear before the text depends on them;
• the level of detail rises and falls naturally;
• the ending resolves the article rather than merely stopping.

This pass is about the arc, not the polish.

Keep the useful roughness

Text often feels machine-written because it is over-compressed and over-polished, not because it contains one forbidden word.

Common symptoms include:

• every paragraph follows statement, explanation, conclusion;
• every list has three perfectly balanced items;
• claims appear as universal principles instead of personal observations;
• sentence lengths and rhythms barely change;
• abstract nouns replace people doing things;
• every section ends by telling the reader what lesson to draw.

Forced fragments, deliberate mistakes, and blacklists of “AI words” do not fix this. They merely create another formula.

Prefer lived observations over declarations. Write “I kept receiving tasks after the important decisions had already been made” before “direct collaboration improves outcomes.” Use concrete verbs. Let a short sentence sit next to a longer one. Sometimes stop before the conclusion and trust the reader.

Natural writing is not unedited. Its editing still serves the person behind it.

Preserve uncertainty

Style editing must not change the strength of a claim.

• “May” must not become “will.”
• A possibility must not become a recommendation without reasoning.
• A rough estimate must not become a commitment.
• An opinion must not become a fact.
• Missing information must not be silently invented.

Semantic drift does not become harmless because it sounds fluent. Review meaning clause by clause, especially in business, technical, legal, medical, and public-facing work.

Writing in another language

Good multilingual writing is rarely a sentence-by-sentence translation.

Languages organize thought differently. Their rhythms change. So do their levels of directness, their ways of addressing the reader, and their ways of moving between a moment and an ongoing process. A translation can preserve every sentence, remain grammatically correct, and still lose the voice.

Treat the task as transcreation:

1. Break the source into facts, intentions, examples, commitments, and desired effects.
2. Put the source text aside.
3. Decide how a native writer would build the same argument for the target audience.
4. Draft in the target language using its own sentence shapes and expressions.
5. Compare meaning, not wording.
6. Ask a native speaker to review tone, register, and cultural assumptions.

The BrIdGe localization method uses a similar sequence: break, ideate, generate. Its researchers found that separating facts from the source phrasing improved fluency without losing meaning. (BrIdGe paper)

Protect names, numbers, technical terms, uncertainty, and commitments. Everything else may need to change: headings, metaphors, examples, paragraph order, even calls to action.

Do not tell a model only to “translate naturally.” Tell it who will read the text, where they live, how the writer relates to them, how formal the language should be, and what the passage should achieve. Then judge the result as writing in its own right.

Choosing a model

There is no best writing model for every task. Creative prose, strict editing, factual synthesis, multilingual work, and high-volume copy each reward different behaviour.

The practical ranking below is dated July 31, 2026. It combines provider documentation, hands-on editorial considerations, and the EQ-Bench Creative Writing v3 leaderboard.

EQ-Bench uses an LLM judge and a fixed set of prompts. Treat its ranking as evidence, not a verdict. Its own methodology makes clear that the scores depend on the judge, the comparison pool, and the rubric.

The providers describe the same differences from their side. OpenAI's GPT-5.6 guidance recommends defining the destination clearly and removing repeated scaffolding. Anthropic positions Opus 5 for complex documents and longer revisions. Google describes Gemini 3.6 Flash as an efficient workhorse for knowledge work and multimodal tasks.

Public leaderboards are useful for English creative writing. For Hungarian, Japanese, Arabic, or any other language, test the models on the language and document type you will actually use. A high English score proves nothing about native rhythm elsewhere.

Once the prompt is stable, a cheaper model may be enough for mechanical transformations. Use a stronger one when the work requires judgment: finding the argument, preserving a delicate voice, reconciling conflicting feedback, or rebuilding a piece in another language.

Ultimately, model choice comes down to personal experience. Writing style is hard to measure objectively, and the model at the top of a leaderboard may not be the one that works best for you.

My personal favourite is DeepSeek V4 Flash. I first noticed it while using it for development. It left more information between the lines, while what it said explicitly felt honest and straightforward. GPT models are excellent at thinking and writing, but they also tend to sugar-coat the reality they describe.

Whatever model you choose, writing this way increases cognitive load. It exposes you to more details, alternatives, and objections. You must judge each one without losing sight of what you wanted to say. The result can be better writing. That does not make it easier writing.