AI Thesis Writing Pitfalls: DeepSeek Warnings and Correct Prompts

Using AI to write a graduation thesis is smooth sailing for some, while others get their work rejected by advisors due to high repetition rates and high AI-generated content rates. A Bilibili content creator recently shared a guide for avoiding pitfalls when using large models like DeepSeek to write theses. The core point is clear: don't let AI generate the entire thesis in one click; instead, use academic prompts to advance step by step. This method involves three layers: literature management, model selection, and prompt design—worth discussing in detail.

Pitfall 1: One-Click Full-Text Generation Leads to High Repetition and AI Rates

The creator starts by pointing out the most fatal move—throwing the thesis topic at DeepSeek^Ad and generating the entire thesis directly. According to them, Chinese repetition rates and AI-generated content rates from domestic AI used this way will be very high, and these two metrics are exactly what current thesis review processes focus on.

Where's the problem? One-click generation means the model, lacking specific literature support, can only rely on generic expressions from training data to piece together content. The resulting sentences have similar structures and stacked terminology, easily colliding with existing literature. Worse, in this generation mode the model often "freely improvises," and there's a lack of real literature correspondence between arguments and evidence—reviewers can spot logical breaks at a glance.

AI-generated content rate (also called AIGC detection rate) is a thesis review metric introduced by universities in recent years to determine the probability that text was generated by a large model. Its principle is usually based on text perplexity and burstiness: AI-generated text often has a uniform probability distribution of word choices and small sentence-length variance, presenting an "overly fluent" characteristic, which detectors use to make judgments. Chinese repetition rate comes from plagiarism detection systems like CNKI, comparing character-level overlap between text and existing literature. The mechanisms differ, but one-click generated full text hits both at once: generic expressions cause collisions with literature, and regular sentence patterns make AI features obvious.

It's worth noting that AI-generated content rate detection is not exact science, and misjudgments occur, but universities generally use it as a preliminary screening reference. The effective way to reduce AI rates is not "rewriting tools" but giving text real literature support and personalized argumentation logic—which is exactly the meaning of the step-by-step process below.

Correct Process: Find Literature First, Then Build a Knowledge Base

The creator's alternative path is a three-stage process: "literature first, outline second, main text last."

Step 1: Use CNKI to Lock Down High-Quality References

Enter thesis keywords into CNKI and select several high-download papers as references. This step seems basic but is the foundation of the entire chain—without real literature, all subsequent AI-generated content is a castle in the air.

Step 2: Let AI Read Literature, Build a Knowledge Base, and Generate an Outline

After downloading the literature, hand it to AI to read and generate a dedicated thesis knowledge base, letting the model thoroughly understand the topic background, then generate an outline based on the knowledge base. The creator emphasizes that the key here is "letting AI thoroughly understand your thesis topic," not skipping understanding and outputting directly. After the outline is confirmed, generate the main text section by section.

The advantage of this step-by-step approach: each section's content is constrained by literature evidence, so the model doesn't dare to improvise; during the outline stage, manual intervention can adjust the structure, avoiding full rewriting.

Thesis knowledge base here refers to organizing downloaded PDF literature and providing it to the model as context. There are two common approaches: first, directly paste or upload literature content into a chat window that supports long context; second, use tools that support file upload or RAG (Retrieval-Augmented Generation) to let the model retrieve literature fragments when answering. The former suits small literature volumes (3-5 papers), the latter suits review-type writing with more literature.

The knowledge base's role is to set "factual boundaries" for the model: when the model is required to argue within specified literature, it reduces reliance on generic expressions in training data, and output is closer to real research context. However, note context length limits—ultra-long literature may cause the model to only "remember" the beginning. In practice, feed in batches by section, or specify in each dialogue round which part of which paper is cited.

Step 3: Use Academic Prompts to Generate Section by Section

The creator mentions a set of thesis prompts summarized by senior students, which can be copied and pasted directly into the dialog box. The video shows usage scenarios of the prompts but doesn't elaborate on specific content.

From a methodological perspective, the value of academic prompts lies in breaking "writing a thesis" into controllable subtasks—such as requiring the model to argue within specified literature, organize paragraphs according to academic norms, and avoid colloquial expressions. The more specific the prompt, the more controllable the model's output.

Model Selection: Why the Creator Doesn't Recommend Doubao and DeepSeek

A more controversial point in the video: recommending ChatGPT over Doubao or DeepSeek for thesis writing. The creator's reason is that the latter two have "very serious hallucinations," and some paragraphs that appear to be model-generated are actually searched content, leading to patchwork theses lacking logic.

This judgment needs dialectical consideration. Retrieval augmentation can indeed bring a sense of patchwork—when the model embeds search results directly into answers, the argumentation chain between paragraphs easily breaks. But ChatGPT also has hallucination issues, just in different forms. A more accurate interpretation of the creator's recommendation should be: choose models good at long-text academic writing with stronger output coherence, rather than simply listing one model as a universal solution.

Additionally, the claim that "Chinese repetition rates and AI-generated content rates are not high" is more of an empirical judgment. Different models perform very differently across disciplines, so it shouldn't be generalized.

Hallucination refers to large models generating plausible but factually incorrect content, typically including fabricating non-existent literature, fictional data, and misquoting authors' views. In academic writing, the harm of hallucination is especially direct: one fake reference can invalidate an entire argument.

Different models do have different hallucination forms. Some models with internet retrieval mix search results with their own generated content; if accepted without distinction, it easily creates a "patchwork feel" lacking argumentative connection between paragraphs. Pure generative models' hallucinations are more reflected in detailed fabrication. Regardless of the model used, the general approach to handling hallucinations is: require the model to annotate the literature source for each argument, then manually verify each against the original text. Making "verification" a fixed step in the process is more effective than relying solely on a "more reliable" model.

Three Practical Suggestions for Thesis Writers

Combining the creator's sharing, several actionable principles can be distilled:

Literature first, refuse to work with empty hands. Any AI-generated content must be anchored in real literature. Download first, read second, generate last—the order cannot be messed up.

Step-by-step generation, retain manual intervention points. The outline and each section's main text must be manually reviewed, and prompts adjusted promptly when problems are found, rather than making major changes after the entire text is generated.

Models are tools, prompts are levers. Rather than struggling over which model to use, focus on prompt design—clarifying literature scope, limiting argumentation angles, and specifying academic style. These are the key variables determining output quality.

At the end of the video, the creator admits these experiences come from years of thesis writing accumulation, with the core appeal being to help everyone avoid pitfalls. For those rushing to finish their graduation thesis, this process is worth trying, but also remember: AI can accelerate writing, not replace thinking. The thesis's arguments, logic, and literature work ultimately still depend on yourself.

One-line Summary

Bilibili creator shares a pitfall-avoidance guide for using AI to write graduation theses: refuse one-click generation, use academic prompts to advance step by step.

Paragraph Summary

Based on a Bilibili creator's sharing, the article outlines the correct method for using large models like DeepSeek to write graduation theses. The core pitfall is throwing the topic at AI for one-click full-text generation, which leads to high Chinese repetition and AI-generated content rates because the model lacks literature support and can only piece together generic expressions. The alternative is a three-stage process: "literature first, outline second, main text last"—first download high-download literature from CNKI, then let AI read and build a knowledge base and generate an outline, and finally use academic prompts to generate the main text section by section. The article also discusses the model selection controversy—the creator recommends ChatGPT over Doubao and DeepSeek because the latter have serious hallucinations that easily lead to patchwork theses, but the author points out this judgment needs dialectical consideration, and prompt design is more critical than model selection.