AI for Writing: Before I Write, After I Write, and What Stays Human

AI can write. That is no longer a particularly interesting observation.

It can write emails, blog posts, product descriptions, headlines, summaries, social posts and stories. It can make them shorter, longer, friendlier, more professional or, if you are feeling particularly adventurous, “more engaging”. Sometimes it does this surprisingly well.

I write professionally. I also write creatively. And I use AI regularly. So the question I find more interesting is not whether AI can write. It is this: Which parts of writing should I actually give to AI?

Because writing is not one activity. Before a sentence appears on a page, someone has to notice something worth writing about. They have to decide what matters, what does not, who the text is for, what it should say and why anybody should care.

Then there is research. Structure. Drafting. Editing. Fact-checking. Rewriting. Deleting the paragraph you spent 40 minutes polishing because it turns out the article is better without it. And yes, eventually there are sentences.

AI is useful in many of these steps. I just don't think it should make all the decisions.

For me, AI-assisted writing works best when AI supports specific parts of the writing process instead of taking over the writing itself. I use it to explore ideas before I write and to edit afterwards. The meaning, judgment, experience and final decisions stay with me.What is AI-assisted writing?

AI-assisted writing simply means using AI during parts of the writing process. That distinction matters. There is a difference between asking AI to write a complete article from a prompt and using it to explore an idea, question an argument, organize research, suggest structures or critique something you have already written. Both technically count as writing with AI. They are very different ways of working. A lot of the conversation around AI writing still starts with the finished text. Can AI write a good article? Can it write a novel? Can it sound human? I think that skips the more useful question: Where in the writing process does AI actually help? For me, the answer is fairly simple. Before writing, AI helps me explore. After writing, it helps me edit. In between, I still want to be the writer.

How I actually use AI when I write

My own workflow is probably less spectacular than asking AI to write an entire article while I make coffee. But it is much more useful to me.

Before I write, AI helps me explore

Often I start with an idea. Not an article, not a headline, sometimes not even a particularly organized thought. Just something I have noticed and think might be interesting. This is where AI can be extremely useful. I can use it to look at the idea from different directions. What questions might people have about it? Is there actual search demand around the subject? Which angle could be useful? What am I overlooking? And, importantly for the kind of work I do, what should this idea actually become? Not every idea needs to become a blog article. Something personal might work better as a Substack essay. A professional observation might belong on LinkedIn. Something that takes 1,500 words to explain but 45 seconds to show could probably be a video. And occasionally an observation about modern work is best explained by a fictional cow.

AI helps me explore those possibilities. It can compare formats, find related questions and show me different ways an idea might work. It can also give me twenty headline suggestions in approximately the time it takes me to decide I dislike nineteen of them. That is useful. But the outcome of this stage is not a finished text. It is a decision. I decide whether the idea is worth pursuing, which part of it matters and where it belongs. Then I sit down and write.

Then I write

This is the part that can get lost in discussions about AI-assisted writing. I actually want to write. Partly because I enjoy it. But also because writing is not simply the mechanical execution of a fully formed thought. Quite often, I discover what I actually think while trying to put it into words. An argument changes while I write it. Something I thought would be important becomes a paragraph I delete. A small observation becomes the centre of the piece. Sometimes the sentence I was trying to reach only appears because the previous three sentences did not quite work.

If AI produces a polished first draft too early, there is a risk that I stop discovering my own thought and start editing somebody else's version of it. That can certainly be faster. I am not sure faster is always the point.

After I write, AI becomes an editor

Once I have written the text, AI becomes useful again. Now I can use it more like an editor or critical reader. Where am I repeating myself? Which paragraph is unclear? Does the structure work? Is there a claim that needs evidence? Does a section answer the question it promises to answer? Is there something missing that would actually help the reader?

AI can also suggest adjustments and alternative formulations. Sometimes they are better than mine. Sometimes they are technically immaculate and make me want to put the original sentence back immediately. Both outcomes are useful. Because this is the part of AI-assisted writing that matters most to me: AI can generate and evaluate possibilities. It cannot take responsibility for my decisions. I still have to decide what stays and what goes. What matters now. What sounds intelligent but is actually nonsense. What genuinely adds something and what merely adds another 400 perfectly acceptable words to a world that is not suffering from a shortage of perfectly acceptable words. And, perhaps most importantly, I have to decide what I want the reader to understand or feel when the text is over.

What AI is good at when you write

AI is very good at making possibilities cheap. Ask for ten alternative headlines and you have ten. Ask for twenty and you have twenty. Ask for fifty and at some point you may begin to regret having invented electricity. The same applies to questions, structures, counterarguments, examples and different ways of explaining something. Writers used to spend considerable time generating possibilities before choosing between them. AI can compress that part of the process dramatically. But generating possibilities and choosing between them are different skills. That distinction runs through almost everything I use AI for.

Brainstorming and exploring ideas

A blank page can be intimidating because it contains no limitations. Anything could go there. AI changes that quickly. I can give it an idea and ask what questions a reader might have, what objections someone could raise, what perspectives I have missed or which related subjects might be worth researching. Suddenly I have something to react to. That last word matters: react.

I rarely think the most valuable function of AI brainstorming is giving me the idea. It is giving me enough material to notice what I agree with, what bores me and what makes me think of something better. AI generates. I select. And selection is already part of writing.

Testing arguments and asking better questions

This may be one of my favourite uses of AI. Writers can become very attached to their own logic. We have spent an hour building an argument, so naturally the argument is excellent. It would be rather inconvenient to discover otherwise. AI has no such emotional investment. I can ask for the strongest counterargument. I can ask which assumptions I am making, which claim needs evidence or what a sceptical reader might question. That does not mean AI is right. It means I now have another position to think against.

For me, that is a much more interesting use of AI than asking it to produce another paragraph. It becomes less of a writer and more of a reader who talks back.

Research assistance

AI can also be extremely useful at the beginning of research. It can help identify terminology, suggest questions, map a subject, organize material and point towards areas that need further investigation. What it cannot do is remove the need to verify information. Large language models can produce statements that sound entirely plausible and are still wrong. They can flatten uncertainty, confuse sources or present an inference with the confidence of someone who has apparently spent the previous 17 years researching precisely this subject.

That confidence is part of what makes them useful. It is also part of what makes them dangerous. For factual writing, I want to know where an important claim comes from. If a statistic matters, I want the original source where possible. If a quotation matters, I check it. If the entire argument collapses when one fact is wrong, “the AI seemed pretty sure” is not a research methodology. AI can help me find the questions. Verification remains my job.

Structure and format

Sometimes I know what I want to say but not yet in which order I want to say it. This is a good AI problem. Give AI an argument, audience, search intent, research and the important questions, and it can suggest several structures remarkably quickly. Again, I would not automatically use the first one.

A logical structure is not necessarily an interesting structure. AI has a tendency to make things beautifully orderly. Introduction. Five sensible sections. Conclusion. Everyone goes home safely. Sometimes that is exactly what a text needs. Sometimes the interesting thought is hiding somewhere AI has politely placed under subsection 4.3. So I use structure generation as comparison. Show me different ways this could work. Then I decide which one actually does.

Editing and revision

There is also an important difference between asking:

What is wrong with this paragraph?

and:

Rewrite this paragraph.

The first gives me information. The second gives me somebody else's sentences.

That can be useful too. If a sentence is genuinely terrible, I am not emotionally committed to preserving its dignity. But the more often I ask AI to rewrite entire passages, the easier it becomes to smooth away the things that made the writing mine in the first place. I prefer to ask where an argument becomes repetitive, which paragraph is unclear, where I am using abstract language, what could be cut or which claim I have not supported. Then I can decide how to fix it. AI becomes an editor rather than a replacement keyboard.

Why does AI writing often sound generic?

This is the strange thing about AI writing. Often, nothing is obviously wrong with it. The grammar works. The structure works. The transitions work. Everything works. And somehow I don't want to read it. Part of the reason is how language models work. They are exceptionally good at producing plausible language based on patterns across enormous amounts of text. Without enough specific material and direction, they tend towards language that is safe and recognizable.

But I think there is another reason.

Good writing is not simply a sequence of good sentences. It contains decisions. What did the writer notice? What did they leave out? Why this example? Why that word? Why begin here? Why end there? Voice is partly language, but it is also judgment.

You can give AI examples of your previous writing. You can teach it that you prefer certain sentence structures, that there are words you dislike and that you would rather not begin every conclusion with “In today's rapidly evolving landscape”. That can help. But I don't think voice can be reduced to a style prompt. You can teach AI that I sometimes write short sentences. Like this. That does not tell it when I believe a short sentence is necessary. The more interesting part of voice happens before those things.

It happens in what I choose to notice.

How do you keep your voice when writing with AI?

I don't really try to make AI sound like me. I would rather make sure I am still doing the parts that make the text mine. That starts before the first sentence. What is the idea? Why does it matter? Where should it live? What do I actually want to say? AI can help me explore those questions. But at some point exploration has to become a decision. That is where I take over. Afterwards, AI can criticize, question, suggest and edit. It can point at the paragraph that seemed profound at 11:30 p.m. and looks considerably less profound the following morning.

But I still make the final call.

Because preserving your voice is not mainly about teaching AI your preferred sentence length or feeding it ten examples of previous articles.

Voice is also what you choose to say. What you don't say. What you find important. What you find ridiculous. What you have experienced. What you notice that somebody else might walk past. AI can imitate patterns in my writing. It cannot have lived the life that created them.

Should you use AI to write a first draft?

Sometimes. This is where universal AI-writing advice becomes difficult. A first draft of a customer email is not the same thing as the first draft of a novel. A product description is not an essay. An SEO article is not a short story. A meeting summary is not a scene in which two characters realize their relationship is over while discussing whether they need more milk. “Writing” contains a lot of very different work.

For functional writing, an AI-generated first draft can save enormous amounts of time. If the objective is to communicate known information clearly, starting with something usable instead of an empty document can make perfect sense. For writing where the thinking itself is part of the value, I am more cautious. Sometimes I don't know exactly what I think until I try to write it. The awkward first paragraph is not always wasted time. It may be where the idea becomes clearer.

If AI gives me a polished argument too early, there is a risk that I begin editing its thought instead of discovering mine. Efficiency is useful. But not every inefficient part of writing is a problem.

AI for content writing vs. creative writing

I make a distinction here. In content writing, the destination is often relatively clear. There is an audience, a question, a business or editorial objective, research and usually some idea of what the reader should understand afterwards. AI can help significantly with the route. Creative writing is different.

When I am working on fiction, not knowing exactly where something is going can be part of the process. A character changes. A scene starts doing something I did not expect. A sentence creates another possibility. The uncertainty is not necessarily an obstacle to remove.

Sometimes it is the work.

That does not mean AI has no place in creative writing. It can help explore possibilities, question a character's motivation, research context, test alternatives or examine a structural problem. But I become much more careful when it starts supplying the actual imaginative decisions.

If I ask AI to decide what my character wants, what happens next, how the scene feels and how the sentences should sound, I may finish the scene faster. I am less sure what exactly I have finished.

What parts of writing should remain human?

I don't think there is a permanent list. AI will get better. Tools will change. Workflows will change with them. But there are things I am reluctant to delegate: what is worth noticing, what I believe, which source I trust, which experience matters, what I find funny, what I am willing to claim and what I am willing to put my name under. And there is something else that is harder to measure.

A text can be grammatically correct, logically structured, factually accurate and still leave absolutely nothing behind.

For me, writing is not simply transferring information from one place to another. I want there to be a thought, a feeling, a message. Something that came from observing something, experiencing something or caring enough about it to sit down and find the right words.

AI can help me work on those words. But the experience behind them is not AI's. It can suggest, question, edit and sometimes make a sentence better. What it cannot know is why something matters to me in the first place, or what I want another person to take away from it.

That part still sits with me. I have to decide what stays, what goes, what is worth saying and what only sounds good. And sometimes that means ignoring a perfectly reasonable AI suggestion because the text would become more correct, perhaps even more polished, but somehow less mine.

So I use AI a lot. Just not for everything. It helps me explore an idea before I start writing and it helps me look at the text again when I am done. The part in between is where I try to figure out what I actually want to say.

And I think I would miss that part if I gave it away.

Frequently Asked Questions About AI for Writing

• Is AI good for writing?

  AI can be very useful for writing, especially for brainstorming, exploring an idea, research assistance, comparing structures and editing an existing draft. I find it less useful when I ask it to make the important decisions for me. The more a text depends on personal experience, judgment or voice, the more I want to stay involved myself.

• How can writers use AI without losing their voice?

  I don't think the best approach is to teach AI to imitate your voice perfectly. I use it to explore ideas before writing and to review my work afterwards, while I write the text and make the final decisions myself. Voice is not only sentence structure or vocabulary. It also comes from what you notice, what you leave out and what you consider worth saying.

• Should I use AI to write a first draft?

  It depends on what you are writing. For functional writing, such as an email or a straightforward summary, an AI-generated first draft can save time. For personal, analytical or creative writing, I am more cautious. Writing the first draft myself is often part of figuring out what I actually think.

• Can AI replace human writers?

  AI can already take over or accelerate many writing tasks, including drafting, summarizing, restructuring and editing. That is different from saying it can replace everything a writer contributes. Experience, judgment, context, responsibility and the decision about what is worth saying still matter. I explore this question separately in Will AI Replace Writers?[internal link once published]

• Can AI help with creative writing?

  Yes, but I would distinguish between supporting creativity and outsourcing it. AI can help explore alternatives, research context, question a character's motivation or identify structural problems. When it starts deciding what a character wants, what happens next and how the scene should feel, I become much more careful. For me, uncertainty is part of creative writing, not always a problem that needs solving.

• What is the best way to use AI for writing?

  There is no single workflow that fits every kind of writing. Mine is simple: I use AI before writing to explore an idea, possible questions, formats and angles. Then I write. Afterwards, I use AI again to challenge, edit and improve the draft. I still decide what stays, what goes and what the text ultimately says.

About Carsten Jan Weichelt

Carsten Jan Weichelt is a content strategist and writer working across content strategy, AI and writing, and creative writing. He uses AI in his professional writing process to explore ideas, research topics and edit his work, while keeping the writing, judgment and final decisions his own.

He also writes fiction and workplace satire, including the observations of a cow called Alma, who has developed surprisingly strong opinions about how humans work.

Connect & Follow him on LinkedIn.