im-not-ai Commentary: Why AI writing Yunmun should design style signals, meaning preservation, and publishing gates before detection scores

1. One-line problem definition

Key line: The real problem with AI writing is not “Was it written by AI?” but that publishers are unable to catch translation errors, missing evidence, or loss of meaning before readers notice them when reading.

As of 2026, blogs, newsletters, landing pages, and internal document drafts are already coming out of tools like ChatGPT, Claude, Gemini, and Codex. So “no use of AI” is not a realistic operational policy. The more important question is Is there a review structure that turns AI-generated drafts into documents that humans can take responsibility for?

This article is for developers and content managers who run Korean technology blogs, corporate content, educational materials, and product documentation. The scope is to use Korean writing style verification skills such as epoko77-ai/im-not-ai as an example, detecting stylistic signals, auditing meaning preservation, verifying sources, and designing gates before publication. Conversely, it does not cover determining authorship of student assignments, gathering legal evidence, or operating AI detection to punish people.

2. First, conclusion

Key one-liners: im-not-ai is more valuable as a “pre-publishing style quality gate” than as an “AI detector”.

The introduction method I recommend is three steps. First, AI tea is not stigmatized with just one score. Second, stylistic signals such as translation, mechanical parallelism, excessive conjunctions, and passive voice are captured in span units. Third, after editing, separately audit whether figures, proper nouns, direct quotations, claims, and sources are preserved.

im-not-ai is a good example in this direction. Lippo deals with translation errors, excessive English quotations, mechanical bullets, “in conclusion/highly significant” idioms, rhythmic uniformity, and excessive formal nouns that are commonly seen in Korean writing written by AI into 10 major categories and over 60 subpatterns. The difference is also specified that Claude Code can be used with a strict pipeline, while Codex and Gemini CLI can be used mainly in Fast mode.

However, it is dangerous to use this tool as “a tool to determine whether a text is human or AI.” The KatFishNet paper also states that Korean AI text detection requires language-specific signals such as spacing, part-of-speech diversity, and comma usage, but detection is a probabilistic signal. In publishing operations, it is safer to replace modifiable quality items rather than detection

3. Decomposition of core structure

Key one-liners: A good markup workflow does not lump detection, rewriting, semantic audit, and naturalness verification.

If we divide the structure of im-not-ai from a practical perspective, it has four layers.

1. Stylist signal detection layer: Look for patterns frequently seen in Korean AI writing, such as “through”, “in”, “first/second/third”, “in conclusion”, and repetition of door-to-door conjunctions. The important thing is that instead of attaching the AI ​​score to the entire document, you select the position to be modified as a span.
2. Surgical margin layer: Fix only detected sections. For example, “Efficiency can be increased through AI technology” can be changed to “AI can increase efficiency.” The writing style should be natural, but the facts should remain the same.
3. Meaning Preservation Audit Layer: Check whether numbers, dates, proper nouns, product names, direct quotes, legal provisions, and key arguments have not changed. Without this step, the statement can easily become a “plausible mistranslation”.
4. Publication judgment layer: Determines whether the final result is natural enough, there are no overstatements, the source link is alive, and the person responsible for the article has finally confirmed it.

In terms of beginner developers, detection is linter, syntax is auto-correction, and semantic audit is testing. Turning on auto-correction without testing can break behavior even if the code compiles. The text is the same.

4. Description of design intent

Key line: Korean AI writing has many signals that are missed when imported directly from English humanizers, so language-specific stylistic rules and meaning preservation rules must be separated.

English-speaking humanizers mainly look at signals such as perplexity, burstiness, and sentence diversity. However, in Korean, signals such as transposition particles, passive voice, long relative clauses, comma position, repetition of formal nouns, and use of English-style pronouns are more directly visible. This is why im-not-ai manages Korean patterns as a separate taxonomy.

The skills structure in the

Claude Code document also fits well with this design. Skills are a way to separate repeatedly pasted procedures, checklists, and multi-step tasks into SKILL.md and load them only when needed. In other words, it turns the sensory request, “Please fix it with good writing,” into a reusable work manual with detection criteria and verification procedures.

An alternative is to give up. The strict pipeline is slower than Fast mode, and review times increase for long articles longer than 8,000 characters. Also, there is no perfect automatic judgment. Rather than handing over publishing responsibility to an AI tool, it should be understood as a structure that makes items to be inspected more clearly by humans.

5. Evidence and Comparison

Key one-liners: Comparisons are made with AI detectors, common text prompts, human editors, and writing style verification skills.

Approach	Advantages	Limit	Recommended situation
General AI detector	Quickly gives probability score	There is a possibility of false positives for each Korean genre and the correction point may not be sufficiently explained	Check general danger signs
General text prompt	Easy to introduce and low cost	Easy to miss numbers, proper nouns, and damage to claims	Short internal memo, low risk document
Human Editor	Contextual judgment and responsibility are most clear	Detecting repetitive patterns and processing large drafts is time consuming	Brand documents, legal/medical/investment documents
im-not-ai style writing style verification skill	Specify Korean AI tea pattern and separate semantic preservation audit	Tool installation and operation rules are needed, and the ultimate person in charge is still a person	Gate before publishing technology blogs, educational content, and corporate knowledge documents

ACL 2025's KatFishNet paper analyzed spacing patterns, part-of-speech diversity, and comma usage for Korean LLM text detection, and reported an AUC-ROC that was 19.78% higher on average than the best existing detection method. These figures can be seen as evidence that Korean requires a Korean-specific signal. However, from an operational perspective, “what to do when a false positive occurs” is more important than detection performance.

6. Actual operation flow / step-by-step execution method

Key one line: The publication gate should not be a decoration that you run once after draft creation, but a checkpoint connected from draft input to final approval.

1. Set the scope of responsibility for the draft.
   Example: “This article uses an AI draft, but the figures, sources, and recommendation judgments are verified by the author.” If it is an internal document, put it in the PR template or CMS checkbox.
2. Fix the style detection criteria.
   Example: S1 is modified unconditionally, S2 is modified when repeated more than 3 times, and S3 is modified only when it overlaps with other patterns. If there are no standards, each reviewer's judgment will be shaky.
3. Divide into Fast and Strict.
   First organize blog drafts of 5,000 characters or less with Fast, and long articles, advertising copy, or externally distributed documents of 8,000 characters or more use Strict or human review. Paste it.
4. Place a thank-you note for preservation of meaning.
   Create separate lists of figures, units, dates, proper nouns, direct quotations, source titles, links, and key arguments and compare before and after modifications.
5. Check the source link immediately before publication.
   Insert references in the main text as HTML links with titles, not as plain text URL listings. Managing clickable links and access dates together makes it easier to catch broken links during CMS migration.
6. Automates the final DoD.
   Example: Includes img, reference anchor format, no plaintext URLs, 7 or more required sections, semantic preservation checked, author perspective included in the publish script. Confirm.

//Document issuing gate example
const gate = {
  styleSignals: { s1: 0, s2Max: 2 },
  fidelity: ['numbers', 'dates', 'names', 'quotes', 'claims', 'links'],
  publishing: ['referencesAsAnchors', 'noPlainUrlList', 'authorViewpoint', 'definitionOfDone']
};

7. Pitfalls

Key line: The most common failures in AI writing style verification are using detection as a punishment tool or automatically applying text without preserving meaning.

• Pitfall 1: Judge people based on their detection score.
  Prevention: Record detection results as “stylistic signals to review” rather than “evidence of AI use”
  Recovery: Reviews that only leave a score Discard and re-extract the editable span and supporting sentences.
• Trip 2: Change the facts while naturally editing.
  Prevention: Put figures, dates, quotes, and product names in the no-edit list.
  Recovery: If the no-edit list changes in the diff before and after the change, Roll back the paragraph and repeat it again.
• Pitfall 3: Mark all posts with the same intensity.
  Prevention: Different allowable change rates for internal memos, blogs, advertising copy, and legal documents.
  Recovery: If it occurs, lower the intensity and remove only the S1 pattern.
• Pit ​​4: Mix source checking with stylistic checking.
  Prevention: Separate stylistic and factual gates. Even if the sentence is natural, if the source is incorrect, publication is prohibited.
  Repair: Rearrange references with primary sources first, and reinforce link titles and dates.

8. Strengths and Limitations

Key line: The strength of the im-not-ai approach is repeatable standards, the limitation is that automation does not replace responsibility.

Strengths are clear. By managing patterns frequently seen in Korean AI articles using taxonomy, reviewers do not have to judge by intuition every time. Using Claude Code's plugin/skills structure, it is easy to share the same procedures on a team basis. The fact that it supports Codex CLI and Gemini CLI Fast mode is also great for use in various work environments.

The limitations are also clear. First, even good people's writing can be flagged for by its uncluttered writing style. Second, correcting the writing style does not improve the basis of the writing. Third, in writing with a strong brand tone, “the way of speaking like a brand” may be more important than “natural Korean.” In this case, rather than applying the taxonomy as is, it should be adjusted with the in-house style guide.

So, I do not recommend using this tool as the sole approval authority. Instead, we recommend having it as a secondary gate through which editors, developers, and marketers pass drafts through the same checklist.

9. Points to study more deeply

Key line: The next step is not to install tools, but to gather failure patterns by document genre in our organization.

• Korean AI text signal: First look at the spacing, part-of-speech diversity, and comma usage analysis of KatFishNet papers.
• Skill operation structure: Look at Claude Code's skills and plugins document and divide individual procedures and team deployment procedures.
• im-not-ai internal structure: View taxonomy, rewriting playbook, and fidelity audit flows. In particular, the principle of “do not touch sections without detection” is important.
• CMS publishing automation: References Automate anchor format, plaintext URL checking, image inclusion, and required section checking with scripts.

10. Action Checklist + Author's Perspective

Key line: The publication standard should be “It is a document for which a person can take responsibility,” not “It does not show any signs of AI.”

• Is there an upper limit to the text strength and change rate for each document genre?
• S1/S2/S3 Are the same stylistic signal standards shared with reviewers?
• Has the meaning of numbers, dates, proper nouns, direct quotations, and core claims been checked for preservation?
• The reference material is in a clickable HTML anchor format, and is there no plaintext URL listing?
• Is there an operating principle that does not use detection results as a basis for punishing people or determining authorship?
• Are the criteria for fast mode and strict/human review documented?
• Has the person responsible for final publication read and approved the text itself, rather than the tool results?

Definition of Done: Stylistic signal S1 0 cases, meaning preservation audit passed, source link confirmed, author's point of view stated, publishing manager approved.

This is my judgment. Teams that write more AI articles should use the term “style verification gate” rather than “remove AI tea.” The former creates a strong sense of hiding, while the latter creates a structure of responsibility. im-not-ai is a good starting point for thinking about its structure in line with Korean document management.

Reference material

• epoko77-ai/im-not-ai GitHub README (Confirmation date: 2026-06-29)
• im-not-ai INSTALL.md: Claude Code, Codex CLI, Gemini CLI Installation method (Confirmation date: 2026-06-29)
• KatFishNet: Detecting LLM-Generated Korean Text through Linguistic Feature Analysis (ACL 2025)
• Claude Code Docs: Extend Claude with skills (Confirmation date: 2026-06-29)
• Claude Code Docs: Create plugins (Confirmation date: 2026-06-29)