用AI进行辅助写作的生产力与创造力

Many people are currently debating whether AI tools should be used to assist writing. In reality, instead of arguing about whether to use them, it is more useful to think about how to use them. I do not reject the use of AI tools in writing, and in practice I use several different ones. Here I am mainly referring to short writing focused on a single topic—such as daily information analysis, news commentary, thematic discussion, or knowledge organization. When it comes to writing academic papers or books, I still do not see AI tools providing substantial help. This may be related to how I understand and position academic writing. At the current level, AI may even suppress systematic and creative writing, because it mainly functions as a tool for linguistic optimization. However, for short, single-topic daily writing, AI can be very useful. So we need to distinguish between the productivity of AI tools and their creativity. Different AI tools also differ significantly in their technical structures, and these differences determine where they fit within the writing process.

The first step of writing is usually the original creation. The task is straightforward—build the framework, expand the concepts, and push the line of thought forward. At this stage I mainly use ChatGPT. Its advantage lies in its ability to handle relatively long contexts. Within a continuous conversation it can repeatedly load previous discussions, allowing the line of thinking to continue without breaking. Technically speaking, each call is stateless, but the system reconstructs the conversation history and feeds it back into the model. As a result, within the same thread it can maintain a relatively continuous progression of ideas.

At a broader level, ChatGPT also retains some long-term information across conversations, such as usage habits or certain structural preferences. This “memory” is managed at the system level rather than being stored inside the model itself, but it can still be helpful when a topic is developed over time. For writing that requires repeated discussion and ongoing revision, this continuity has real value.

Once a draft is produced, the second step begins—local logical calibration. At this stage the goal is no longer expansion but clarity, precision, and rigor. I usually use Google Gemini for this kind of work. Gemini’s design places more emphasis on the task itself and on clear conversation boundaries. In enterprise environments it generally does not automatically integrate historical context across conversations; each thread remains relatively independent. During the creative phase this may feel fragmented, but during calibration it becomes an advantage. It behaves more like a tool that reexamines a problem from scratch, without being influenced by the earlier path of expression. This makes it well suited for checking whether a particular line of reasoning holds or whether a concept has been expressed clearly.

In simple terms, Gemini emphasizes the independence of each conversation, while ChatGPT emphasizes the continuity of dialogue. This structural difference corresponds well to the different needs of the generation stage and the calibration stage.

The third step is global structural integration. The absence of obvious problems in individual sections does not guarantee that the entire book is free of structural risks. At this stage the focus shifts to whether core concepts remain consistent across different chapters, whether terminology changes over time, and whether there are repetitions or contradictions. For this task I use NotebookLM. NotebookLM operates through a RAG mechanism—document-level semantic retrieval. It does not rely on chronological chat history. Instead, it vectorizes the entire manuscript and retrieves relevant passages from the whole document when a question is asked. As long as the source material remains unchanged, any number of new conversations will still analyze the same underlying text.

In other words, NotebookLM’s cross-conversation continuity comes from the document itself rather than from chat history. As long as the material is stable, the structure remains stable. This makes it particularly suitable for checking overall consistency.

The division of labor among these three tools is quite clear. ChatGPT relies on conversational context and a limited form of long-term information, which makes it suitable for generating and developing ideas. Gemini emphasizes conversational independence and is better suited for local logical analysis. NotebookLM relies on document retrieval and is therefore more appropriate for scanning the structure of an entire book or long article. In practice one can move back and forth between the three—first generate, then calibrate, then scan the overall structure, and finally return to the generation stage for revision.

Even when adopting this collaborative approach, it is essential to retain the ability to complete a work independently without AI. Technology solves problems of efficiency and stability; it does not replace thinking itself. A writer should be able to start from zero, construct the framework, push the reasoning forward, and write structurally complete sections without any assistance. If that ability disappears, multi-tool collaboration easily turns into dependence.

This article itself was produced through direct voice dictation. The overall structure, sequence of ideas, conceptual hierarchy, and argumentative logic were all developed in a single pass during spoken expression. Only afterward was AI used for minor adjustments at the language level rather than for constructing the content itself.

Voice writing requires a particular skill. When a sentence is halfway spoken and the angle of expression turns out to be inaccurate, one cannot simply stop and restart from the beginning. Speech is a continuous output; repeated backtracking breaks the flow of thought. A more suitable method is to correct the sentence internally while continuing to speak. For example, if one initially intends to use an active construction but realizes that the logical emphasis is misplaced, it can be converted into a passive form on the spot. Or if a condition appears insufficient, a qualifying clause can be added so that the earlier part of the sentence extends naturally rather than being replaced. When the sentence finally lands, it still remains smooth, coherent, and grammatically sound.

This ability itself requires training. Linear language output must be supported by a three-dimensional structure of thought behind it. The line of reasoning must already be largely formed before expression begins, or at least the direction must be clear. Otherwise, once speech starts, the result quickly becomes a patchwork of improvised corrections.

The purpose here is not to instruct AI how to write, nor to provide prompts for it to expand content. What we present is an expression that has already been formed through spoken reasoning. AI merely performs minor optimization at the level of language rather than replacing the generation of thought. If the expression itself is chaotic, no tool can truly repair it. On the contrary, the more mature the expression, the more the tool’s role becomes fine-tuning rather than reconstruction.

Basic ability must be maintained over the long term. Just as a pianist practices technical fundamentals every day—even when performing complex works on stage—the underlying control comes from repeated training. Writing is similar. The ability to think, judge structure, and organize language must still function independently without tools. Otherwise, the stronger the tools become, the more human ability deteriorates.

Under this premise, using AI is a rational choice. Rejecting technology does not guarantee originality. The real question is whether the author retains control. Original ideas come from the writer, the framework is designed by the writer, the logic is handled by the writer, and the final judgment is also made by the writer. AI can accelerate generation, offer alternative perspectives, and help calibrate expression, but it does not determine direction.

As long as the direction remains in the writer’s hands, AI remains a tool. The subject of writing is still the human being. The significance of multi-tool collaboration lies in improving efficiency and structural stability, not in outsourcing thought. Within this structure, we are using AI rather than AI using us.

One more point needs to be added. The process described above mainly applies to knowledge-oriented writing built around a specific theme—such as blog essays, information analysis, or periodic knowledge organization.

In this kind of writing there is also a practical question to consider: who is the content written for?

In the past the assumed reader was human. Now the situation has changed. In many cases the first reader of an article may not be a person but a machine. When people encounter a text, they often place it into an AI tool first to extract a summary or condensed version, and only afterward decide whether to read the full article. The text may pass through an algorithmic filter before it reaches human judgment.

Therefore, writing strategies cannot focus only on human reading experience. For informational or analytical content, the first reader may very well be an AI system. The logic must be clear, the concepts stable, the terminology consistent, and the structure explicit. Otherwise the text may be compressed or even misinterpreted during the machine-processing stage.

For emotional or artistic expression, however, the primary reader is still human. Such writing emphasizes rhythm and experience, and it does not need to prioritize algorithmic summarization.

Writing therefore needs to distinguish between different potential reader structures. Some content must consider machine interpretability, while other content should focus on human comprehension. The coexistence of these two kinds of readers is the reality today.

This means writing is no longer only communication between people; it also involves interaction between humans and algorithms. Tools participate in writing, and they also participate in reading. We use AI to assist expression, while at the same time facing AI as a filter.

In this environment the essential issue remains control. The writing environment is changing, but the subject of writing has not. The task is to continue holding the direction within this new structure rather than being led by the tools.