2026 Update: What to Do When a Paper's AI Rate Is Too High? A Guide to Handling AIGC Suspicion Rates

The AIGC Suspicion Rate Is Too High!

#

During graduation season, the easiest thing to get stuck on is not necessarily plagiarism checking, but the AIGC suspicion rate. Plagiarism checks at least show repeated sources. AIGC detection is more like a probability judgment: it only tells you whether a paragraph looks like it was written by AI, but does not clearly explain the basis for that judgment or tell you how to revise it.

These results are indeed controversial. Related CCTV reporting mentioned that many students said their own writing was also counted as AI-generated, and experts suggested that AIGC detection should not be used as the only evaluation basis¹. The People’s Daily-affiliated Satire and Humor also reported cases where Moonlight over the Lotus Pond was judged to have a 62.88% AI suspicion rate, and Preface to the Pavilion of Prince Teng was detected as having a 100% AI rate.²³

But when schools have hard requirements, we still need to deal with them. This article uses Gecida as an example and explains a steadier process: not turning the paper into “casual chatter,” and not evading academic review, but using the report to locate high-risk paragraphs and changing overly templated, vague, mechanical expressions back into writing that is more concrete, verifiable, and consistent with your own research process.

I tested one draft myself, and the Gecida AIGC suspicion rate dropped from 46% to 16%. This result only represents that draft and that detection environment. It does not guarantee the same number can be reproduced for every paper, but the process itself is relatively stable.

The full revision-method document can be downloaded from Quark Netdisk, or viewed in the GitHub repository.

1
2
3
4

I shared “降低AIGC疑似率260526” with you via Quark Netdisk. Click the link or copy the full text, then open the “Quark APP” to get it.
动作昳张凤凰台多好
/~47b93YkcoV~:/
Link: https://pan.quark.cn/s/031285afb384

Method 1: Use OpenCode + DeepSeek-V4-Pro for Segmented Revision

#

This is the method I recommend most, suitable for people who can already use command-line or Agent tools.

I personally recommend the OpenCode + DeepSeek-V4-Pro combination: strong model capability, low price, very long context, a relatively smooth connection flow, and China-friendly access. In my test, revising one paper cost around one yuan, with some variation depending on article length.

1. Prepare Three Files

#

Put the following three files in the same directory:

1
2
3

risk-report.pdf        # Gecida AIGC risk-rate report
paper.docx             # original paper
aigc-rewrite-guide.md  # revision-method document

If the report is a webpage screenshot, you can export it as a PDF or organize it as Markdown. The point is to let the Agent see which paragraphs were marked as risky, what the original text is, and what the revision rules are.

Do not only throw the high-risk paragraphs to it in isolation. Many sentences can be revised when viewed alone, but may disconnect from the surrounding text when placed back into the paper. So it is best for the Agent to read both the original paper and the detection report.

2. Install and Connect the Model

#

If OpenCode is not installed yet, first visit the OpenCode official website to check the installation method.

After installing OpenCode, run the following in the paper directory:

1

opencode

After entering OpenCode, type:

1

/connect

Select deepseek, enter the DeepSeek API Key, then select DeepSeek-V4-Pro. If you do not see it in the model list, run:

1

/models

There is a small pitfall here: if the model is unavailable, first check whether the API Key is valid, whether the balance is sufficient, and whether the model name is selected correctly. OpenCode’s troubleshooting documentation also reminds users that when a model is unavailable, they usually need to first check whether the provider is authenticated, whether the model name is correct, and whether extra permissions are required.

3. Use This Prompt

#

Send the following directly to the Agent:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

risk-report.pdf is my Gecida AIGC risk-rate report, paper.docx is my paper, and aigc-rewrite-guide.md is the revision method.

Please make full use of subagents to complete the following tasks:

1. First parse risk-report.pdf, locate medium-risk and high-risk paragraphs, and list the corresponding original text.
2. Then read paper.docx and confirm the context of these paragraphs in the full paper, avoiding revisions detached from context.
3. Based on the rules in aigc-rewrite-guide.md, provide revision suggestions only for the sentences and paragraphs that need to be revised.
4. During revision, you must preserve the original meaning, technical terminology, data, citation relationships, and research conclusions. Do not fabricate experiments, cases, interviews, data, or references.
5. After each revision, self-check whether any of the following problems exist:
   - Facts were changed;
   - Terms were replaced incorrectly;
   - Sentences became colloquial but no longer fit academic style;
   - Citations, data, or section numbers were lost;
   - The revised text no longer connects naturally with the context.
6. If the self-check fails, revise again.

The final output should be a Markdown document. Output format:

- Output by natural paragraph;
- Each item includes “Original,” “Revised,” and “Revision note”;
- Both “Original” and “Revised” should be wrapped in separate code blocks;
- If only part of a source paragraph needs revision, split it into multiple items;

If you do not have a detection report yet, delete the part about risk-report.pdf in the first sentence:

1

paper.docx is my paper, and aigc-rewrite-guide.md is the revision method.

You can also ask the AI Agent to revise the whole text and output a report at the end.

Method 2: If You Do Not Use Agents, Process Segments with a Web-Based AI

#

If you are not familiar with OpenCode, API Keys, or the command line, you can directly use a web-based AI.

This method is simpler, but its context management is weaker. So you must process the paper in segments and should not paste the whole paper at once. Handling 3 to 5 natural paragraphs each time is enough. When there are too many paragraphs, web-based AI easily forgets the rules set earlier.

First send the revision-method document to the web-based AI:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

Please first learn the content of this document and explain in bullet points which revision principles you learned.

Next, I will send paragraphs from my paper with a relatively high AIGC suspicion rate. Please revise them according to this document, but you must follow these requirements:

1. Do not change the original meaning;
2. Do not change technical terminology;
3. Do not change data, conclusions, citations, or section numbers;
4. Do not fabricate any research process, cases, experiments, interviews, or references;
5. Output “Original / Revised / Revision note”;
6. The revision goal is to make the expression more concrete and more consistent with my research context, not simply more colloquial.

After it restates the rules, send the high-risk paragraphs:

1
2
3

Below are the paragraphs that need revision. Please output according to the rules above.

Original: xxxx

Then paste the original text.

This method is suitable for handling 3 to 5 natural paragraphs at a time. When there are too many paragraphs, web-based AI easily loses the rules, and later content may become more and more generic.

There is also a small pitfall here: do not ask it to simply execute instructions such as “reduce logical connectors” or “use fewer technical terms.” The suspicion rate may drop in the short term, but the paper will get worse.

Method 3: Use Skills to Solidify the Revision Rules

#

If you often use AI Agents, you can use this skill I organized.

1
2
3
4

I shared “降低AIGC疑似率260526” with you via Quark Netdisk. Click the link or copy the full text, then open the “Quark APP” to get it.
动作昳张凤凰台多好
/~47b93YkcoV~:/
Link: https://pan.quark.cn/s/031285afb384

GitHub link: aili-notes/skills/aigc-paper-rewrite at main · Rosetears520/aili-notes

Tell the Agent directly:

1

Help me install this skill: https://github.com/Rosetears520/aili-notes/tree/main/skills/aigc-paper-rewrite

After starting OpenCode, prompt it to load this skill:

1
2
3
4
5

And make full use of subagents to complete the following tasks:

1. First parse risk-report.pdf, locate medium-risk and high-risk paragraphs, and list the corresponding original text.
2. Then read paper.docx and confirm the context of these paragraphs in the full paper, avoiding revisions detached from context.
3. Let the subagent use aigc-paper-rewrite to revise and output a report.

The principle of this method is similar to the first two. The difference is that you do not need to copy and paste the revision rules every time, so it is suitable for repeatedly handling multiple documents or multiple revision rounds.

But I have not run very strict batch testing on this method yet, so I recommend treating it as an auxiliary workflow. Whether to replace the text in the end should still be reviewed item by item by yourself.

The Final Step Shared by All Methods

#

No matter which method you use, after the revision is complete, remember to revise it again yourself!!!!

At minimum, check these issues:

1. Whether technical terms were randomly replaced;
2. Whether data, years, citations, and section numbers changed;
3. Whether sentences became too colloquial just to “sound human”;
4. Whether the context still connects naturally;
5. Whether the revised expression still matches your own research content.

Especially for technical terms, do not casually replace them just to lower the suspicion rate. Terms such as Transformer, principal component analysis, difference-in-differences, and F1-score should remain what they are. The detection rate is not the only goal; the accuracy of the paper is the bottom line.

Another important point: if the paper contains enterprise data, unpublished experimental results, or supervisor project materials, do not upload them directly to uncertain platforms. You can desensitize first: remove company names, personal names, IDs, and unpublished data, leaving only the text structure that needs revision.

Avoid this situation: some students, in order to reduce the AI rate, had to delete logical words and add casual filler, causing a sharp drop in paper quality⁴. That kind of situation can become embarrassing!!!

References

#