I Used GPT to Learn How to Write Less Like GPT

Introduction

I usually wrote by assembling blocks of thought.

In the past, writing itself took about 20% of the time, while revision took almost 80%. Writing in a way that people could easily understand was not an easy task for me.

Back then, however, GPT connected those blocks smoothly.

I would tell GPT my thoughts without holding anything back, and GPT would respond with a degree of sympathy. Based on the conversations that accumulated through that process, I was able to expand my thoughts further, and I believed I had become better at expressing them more precisely.

Of course, I still revised the text. Or rather, it would be more accurate to say that I learned a better way of writing through revision. By looking at the smoother writing GPT produced, I also learned how to write. Then I added what the AI had missed, or what I personally wanted to add. It was a kind of finishing touch.

The process was something like this: GPT would smooth out my jagged lines of thought into something more coherent and rounded, and then I would insert sharper sections back into it.

The Process

I usually wrote through the following sequence of steps.

1. Free conversation

I choose a topic and unpack my thoughts through conversation with AI. Sometimes this stage is loose and disorganized. The topic often begins from an experience, or from a sudden flash of concern.

At this stage, I am not trying to find the correct answer. I am trying to secure useful expressions.

2. Excerpting

From the entire conversation, I strengthen “my own sentences.” I treat AI-generated sentences only as references, and place my own voice at the center.

3. Structuring

I rearrange the excerpts under subheadings, organize the flow of the argument, and remove repetitive sentences.

4. Verification

When necessary, I check the facts. I separate the realm of opinion from the realm of fact and review them accordingly.

5. Revision

After reading the smoother version connected by AI two or three times, I remove unnecessary sentences, add new ones, and insert necessary sources, images, or other materials in between.

Revision

The biggest obstacle is the distinctive expression associated with AI, or what I call the “GPT style.”

• Contrastive declarations such as “This is not A, but B”
• Em dashes, bold text, and exaggerated emphasis
• Parallel renaming through abstract nouns
• Habitual endings in the form of questions

Part of my resistance to GPT style comes from the fact that it feels unpleasant. But another reason is that it is very different from my own way of speaking and writing. Expressions such as “layers,” “another name for,” or “translated into the language of…” are especially grating to me, as are question-style endings. I am not saying I never use such devices at all, but they are overused to an absurd degree. I do not write like that. If they seem unnecessary, I remove them during revision.

I also had to change punctuation marks I do not normally use in my writing.

I would replace an em dash ( — ) with a hyphen (-), a middle dot (·) with a comma, and an arrow (→) with a hyphen and angle bracket (->).

There is also the issue of repetition. But beyond that, there is my aversion to GPT style itself. The source of that aversion may be a resistance to generated text that does not feel sincere, or perhaps an unconscious resistance to AI.

In principle, even if the expression is not wrong in context, and even if there is no problem with conveying the content, it can still become a problem because of its impression. Once a way of speaking develops a certain atmosphere around it, it can become marked as something to avoid.

For example, in the Korean military, there used to be a way of ending sentences with a particular stiff military phrase. It was common in the 2000s, and perhaps even into the early 2010s. At least in the unit where I served, that manner of speaking had already fallen out of use. If someone spoke that way, they would often be teased or corrected.

This is only my personal opinion, but I suspect that military sitcoms from the early 2010s, such as Blue Tower, may have contributed to its disappearance by adopting that speech pattern. I am not saying that was wrong. I simply think that a way of speaking that once felt normal can change because of atmosphere and impression. Eventually people begin to feel resistance to it, and in the end, it disappears.

When I use contrast, I use various methods. I do not consciously set out to use those methods. They simply appear naturally as I write. This is not something that happened because of GPT. I had always written that way.

There is the GPT-like method of saying:

“Not A. Not B either. Rather, C.”

But there are other ways.

1. Simply say, “It is C.”
2. If potential objections or qualifications are important, they can be developed in the following paragraph. For example: “Of course, one might think of this as A or B. But that is not what I mean. For instance…”
3. When there is a paragraph that states “It is C,” one can add a parenthetical note, or, in the case of a book, attach a footnote.

AI systems like GPT repeat the “Not A. Not B either. Rather, C” structure far too often, almost obsessively. I have often had to struggle through revision because of it. So when I notice this kind of pathological repetition, I keep adding it to memory: do not do this. But it does not listen very well.

Attitude

I am not trying to teach other people how I write, or present this as a special method.

This is not a method that applies to everyone. In fact, it may be quite an exhausting way to write. I do think AI can organize context, find sources, and fix illogical sections, contradictions, and disorganized parts. But the content itself should not be written by AI.

The act of presenting an opinion, making an argument, and saying something one truly thinks belongs to the human being.

AI merely retrieves the most desirable and probable responses from what exists within its data. No matter how much one tries to control it through prompting, how frustrating and dull is that?

A 2,000-character response comes out.

“This is not it. Write it again from the beginning.”

Another 2,000-character response comes out.

“This is not it. Write it again from the beginning.”

Another 2,000-character response comes out.

“This is not it. Write it again from the beginning.”

AI should not be used like Doraemon.

Is that even fun?

Does it not leave a hollow feeling somewhere in the chest?

Does it not feel heavy and suffocating?

For me, it does.

Closing

There may be situations where one has no choice but to borrow the power of AI for an exhausting task, even if reluctantly. But if that is not the case, I feel uneasy about entrusting a tool with the task of writing out my thoughts entirely.

As much as possible, I want to write my own work.

If I spend all day making GPT write for me, the result may become nothing more than a tedious, stale, overly academic piece of writing that wastes the reader’s time.

After using the process described above for about three months, I became much faster at writing. I also became able to write with a much more organized sense of context. It is genuinely a good thing that I now rely on GPT far less than before.

GPT is good at organizing context. But it often produces writing with a very particular style. It softens parts that should be sharp, blurs what should be said directly, and creates text in which my own opinion is not fully reflected.