Make Korean AI-Written Text Sound Natural Without Changing Its Meaning

## Keep the Original Facts and Tone, and Edit Only the Problem Spots Treat humanization as a constrained edit, not an opportunity to reinterpret the text. Preserve every fact, claim, number, proper name, quotation, and source-specific commitment. - Keep the original genre: do not turn a report into an essay or a column into literary prose. - Preserve register in both directions. Formal Korean must remain formal, and conversational Korean must remain conversational. - Do not replace natural spoken endings with stiffer forms merely to sound polished. - Change only spans connected to a detected problem. - Remove AI-like signals without introducing new clichés, facts, examples, or rhetorical claims. - Leave already-natural passages alone. Fidelity outranks elegance. If a natural-sounding rewrite would alter meaning, keep the more faithful wording. ## Find AI-Like Patterns and Sort Them Into Three Levels of Seriousness Diagnose observable spans before rewriting. Classify recurring signals such as translationese, excessive English quotation, mechanical parallelism, canned phrasing, passive-voice overuse, connector overload, uniform rhythm, and excessive emoji or bullets. Use three severity levels: - **S1 — decisive:** one occurrence strongly signals machine-written text and should be removed. - **S2 — strong:** tolerate one or two isolated instances, but revise repeated occurrences. - **S3 — weak:** act only when the signal overlaps with other patterns. For ordinary and intensive runs, identify the three to six dominant patterns across the whole draft. A diagnosis is evidence for targeted editing, not permission to rewrite unrelated passages. ## Choose a Light, Standard, or Heavy Edit Based on the Draft Derive a quantitative route recommendation from the draft, then let an explicit user choice override the default. - **Light:** one conservative rewrite for a strong draft with almost no lexical AI signals. If there is little to fix, stop early and say the draft is already sound. - **Standard:** one diagnosis followed by one targeted rewrite for a typical AI draft. - **Heavy:** diagnosis, rewrite, deterministic change-rate verification, and final source comparison for dense problems, texts over 15,000 characters, or work that requires validation evidence. A strict or precision request forces the heavy route; an explicit request for a light pass forces the light route. If metrics fail, degrade gracefully to the standard route instead of pretending that routing succeeded. When a light or standard result grades poorly, recommend a heavy rerun. Do not escalate automatically without the user's choice. ## Run Only the Needed Steps, Adding a Final Check Against the Source for Heavy Edits Keep each pass responsible for one decision. 1. Prepare the source and calculate routing signals. 2. For standard or heavy work, diagnose the dominant patterns across the full text. 3. Rewrite only the grounded spans, applying conservative intensity for light work and targeted intensity for standard or heavy work. 4. Apply the deterministic change-rate gate on every route. 5. On the heavy route, compare the rewrite directly with the source, check fidelity and naturalness, and correct only the remaining problem spans. Light work uses one rewrite pass. Standard work normally ends after diagnosis and targeted rewriting. Heavy work adds final source-based review; it must not become an excuse for a broad stylistic rewrite. ## Split Long Texts Only When a Heavy Edit Needs It Prefer a single rewrite call because repeated chunks reload rules and diagnostic context, increasing cost and inconsistency. Do not recommend chunking for texts of 15,000 characters or fewer. Chunk only on the heavy route, and only when preparation actually produces two or more chunks. If preparation produces one chunk, use a single call. When chunking is necessary: - preserve untouched source passages as passthrough text; - process independent chunks without sharing mutable result files; - reassemble them in the original order; - compare character counts and boundaries after reassembly; - run the same change-rate and fidelity gates on the complete result. Efficiency comes primarily from reducing unnecessary model calls, not from silently lowering model quality. ## Keep Natural Spoken Korean Endings Instead of Making Them Stiff Treat AI-like style as a problem of grammar, rhetoric, repetition, and rhythm—not as proof that the text is too informal. - Preserve conversational endings and idiomatic turns when they are natural in context. - Do not mechanically replace common contracted forms with more formal alternatives. - Keep formal input formal without making it ceremonious. - Vary rhythm only where uniformity is a diagnosed signal. - Remove stock praise and bureaucratic phrases only when they are present; never insert new ones. - Keep quoted text unchanged. The goal is a human voice consistent with the source, not a single preferred tone. ## Measure How Much the Text Changed and Stop When the Edit Goes Too Far Measure the edit rather than relying on the model's impression of convergence. - Warn when the change rate exceeds 30 percent. - Stop the workflow when it exceeds 50 percent. - In heavy review, compare the source and rewrite against a comprehensive fidelity checklist covering meaning, facts, quantities, names, quotations, genre, register, omissions, and additions. - Check naturalness in both directions: remaining AI signals and over-editing. - Correct only the spans that fail review. If the final review cannot safely resolve a conflict, hold the result and report the issue for human review. A deterministic change-rate gate decides whether the edit is within bounds; repeated model self-diagnosis does not replace it. ## Grade the Result From A to D and Say What to Do Next Use severity counts, improvement, and over-editing signals to classify the result: - **A:** no S1 findings, at most two S2 findings, and at least 70 percent score improvement. - **B:** no S1 findings, at most four S2 findings, and at least 50 percent score improvement. - **C:** one or two S1 findings, or two over-editing signals; recommend a second pass. - **D:** three or more S1 findings, or severe over-editing; recommend human review. Maintain one authoritative taxonomy with stable identifiers. Add or promote patterns only through reviewed evidence, regenerate derived quick rules from that source, and use regression fixtures plus deterministic checks to prevent drift. Report the chosen route, material edits, gate results, unresolved risks, and whether another pass or human review is recommended.