[行业应用] Why is the content produced by AI unsatisfactory and unlike something written by a human, and how can this be changed?

[baike]Why Doesn’t It Seem Human[/baike]
Because ChatGPT itself is a language application built on OpenAI’s base model. Its design is for content processing, including article generation, text analysis, data analysis, and a range of other text-related functions. Facing a huge user base, to achieve general applicability, they set some defaults—for example, article processing uses a default style, layout, vocabulary, etc. For those of us who professionally produce content, this is clearly not enough. But it also supports a certain degree of customization, and through prompt engineering you can develop it as much as possible according to your style. If you don’t understand these things well, it will be very difficult for you to set standards that meet expectations.

Another reason is that ChatGPT’s base model has a limited knowledge base. Their knowledge base is collected and purchased, and most of it is public data. It learns from this data to generate content for us. Its understanding of industry knowledge and professional knowledge is limited, because much professional knowledge is closely guarded by us—it is hidden in our minds. We will not easily disclose these industry secrets. Therefore, when it comes to professional fields, it knows very little, because it cannot write professional knowledge for us, nor can it provide us with relevant case citations, application methods, etc.

Based on the above two main reasons, it is very difficult to directly use ChatGPT to generate articles that meet expectations. Next, the moderator will teach everyone the methods. Please read patiently. Among our trainees, most are already using this method to replace writing, and the results are very good.

[baike]Specific Manifestations[/baike]Before understanding the methods, let’s first understand the specific ways ChatGPT does not seem human, including: no emotion, overly objective and cautious, does not make excessive inferences, does not generate thoughts, lacks subjective guidance and interactive elements, does not appear as a human persona, is not professional enough, has a single vocabulary pool, cannot incorporate current events, etc. These combined manifestations make it not seem like a person.

[baike]Two Learning Methods[/baike]
1. ChatGPT Prompt Engineering
Through the above shortcomings, we can fill in the missing parts. This is a clumsy method, but it can improve a lot. Next, use prompts to improve it as much as possible, add your own personalized settings at each stage, and finally tell it to increase a certain amount of randomness—that is, temperature. Treat the entire writing task as a systematic project. We need to break down the composition of the article step by step, which requires you to understand writing well enough.

Your prompt engineering will roughly be as follows:

1. Please create an article titled XXX. My purpose is XXXX, my readers are XXXX, and I plan to post it on the XXX platform. Now please provide several styles for the article.
2. Okay, I choose the XXX style. Now please list several article structures and outlines.
3. I want to use the XXX structure and outline.
The above is a very good start. Now I will combine it with the knowledge base method to make it fill in the missing parts.
4. I have uploaded these documents; they are my writing content. I hope you learn my writing methods, word usage, etc. Since this is the first time writing, please analyze it as a whole and paragraph by paragraph. First tell me how you understand my writing. I hope you will write as closely as possible in that way.
5. Okay, I’ll add a few points XXXXX.
6. I have uploaded these documents. Please take a look—this is my persona content. Please carefully study what kind of person I am and what my experiences are.
7. Okay, next please learn my persona, imitate me and the writing samples I provided you earlier, and let’s start creating an article, XXXXX.

2. Build an Exclusive Large Model
Through the above methods, we have basically been able to learn how the moderator combined an understanding of the ChatGPT product with an understanding of article writing to design prompt engineering. You will get a writing machine that leans toward your expected development, but in reality it is still far from enough and not intelligent enough. Next, the method the moderator tells you is the most perfect solution.

That is to build your own large model. Everyone, don’t get a headache—this is actually not that difficult. If you have studied our knowledge and tutorials, you can get it done in a few days. If you haven’t studied them, you can still learn it by slowly exploring on your own for a month. Building an exclusive large model means fine-tuning on OpenAI’s base model, incorporating our writing content and requirements so it fully learns, and ultimately forming an exclusive model. From then on, you only need to tell it in one sentence what your article title is, and the corresponding article will come out.

Methods for Fine-Tuning a Large Model
1. Prepare your sample articles and the style you want to write in—the more, the better.
2. Organize your articles into the format required for fine-tuning a large model, along with the corresponding system instructions.
3. Purchase an API.
4. Upload your data format to the fine-tuning large model and start training.
5. Use scripts to call the API and the large model, and start writing articles.

If you are a beginner, this method may be a bit difficult for you to understand. This actually proves that you do not yet understand OpenAI’s product line and what they are doing. Simple prompt engineering has not even reached the entry level. Truly, at a later stage, we will definitely learn automation and how to build various exclusive large models. Only these methods can truly solve practical problems and meet enterprise-level requirements.

Here we can simply understand it as turning your articles into a knowledge base for the large model to learn, and then incorporating it. It learns all your writing characteristics, much richer than the elements described in our prompt engineering—for example, your exclusive vocabulary, writing methods, integration of current events, and your industry knowledge. This model belongs only to you, and it will be based mainly on your knowledge. You can also set its basic output methods. Now that you have it, you only need to input an article title, and an article almost identical to your writing style will instantly appear.

How can you obtain good article examples? If you are not an expert writer, then you can find people in this industry who truly write well, get their content, organize it, and use it for training.


Now, it is impossible for you to master how to use a fine-tuned large model for articles through just these few paragraphs from me, but it is your only answer. Everyone can learn through our large model fine-tuning section. If you want to learn faster, you can contact the moderator for 1-on-1 private tutoring and learn our overall systematic ChatGPT teaching.