A Korean AI humanizer, examined

Worth stating precisely, because this category is full of pages describing a proprietary Korean pipeline that does not exist. A request leaves the browser carrying four things: the route, the engine family, your Korean text, and an optional instruction string. On the server a language model receives a single directive, to rewrite the supplied text to accomplish the transformation the tool name implies, to preserve the original meaning and factual claims, and to return only the rewritten text with no preamble, commentary or fencing. That is the entire mechanism.

There is no Korean corpus behind it, no morphological analyser, no honorific rule table and no idiom database. There is a capable model reading Korean and writing Korean, constrained to hold your claims.

Architecturally that makes this a remote route, not a setting that could be flipped. Text is transmitted for the run, input is accepted only between 1 and 12,000 characters, and the allowance is 5 free runs per rolling 24 hours keyed on a hash of the caller's address. The deterministic cleaners elsewhere on this site transmit nothing at all.

Four patterns are worth isolating, because they fail differently. Two are lexical and a rewrite reliably moves them. Two are structural and it often does not.

해요체 as a default

A register chosen by the model rather than by the document's purpose, then held without variation. Lexical in appearance, but it is a decision about audience, which is why an unsteered rewrite tends to reproduce it.

Absent discourse markers

글쎄, 아무래도 and 그러니까 carry hedging and consequence. Their absence flattens the modality of every sentence to plain assertion. A rewrite moves this readily, because it is word choice.

Connective monoculture

그리고, 하지만 and 따라서 doing every join in a long piece. Also lexical, also readily moved, and the easiest of the four to audit: count them before and after.

Uniform paragraph geometry

Every paragraph the same length, opening on the same construction, closing on the same one. Structural, and the pattern most likely to survive intact, because the instruction constrains meaning rather than shape.

Segment first. Rewrite fidelity degrades as input length and topic spread rise, and length is the single controllable variable in the run.

Step 1

Paste Korean between 1 and 12,000 characters. The bound is enforced before the request leaves, so an over-length draft is refused rather than silently truncated at the boundary.

Step 2

Specify the register in the instruction field, up to 300 characters, appended to the system prompt. Unspecified register is not neutral: it is delegated.

Step 3

Diff the output. Sentence endings first, for speech-level consistency, then every clause carrying a number, a name or a date, because meaning preservation is an instruction to a model and not an assertion by a parser.

Every commercial page on this term sells speech-level handling as engineered behaviour: consistent 존댓말 endings, no awkward mixing of levels, honorifics matched to context. It is the standard claim of the category.

The instruction this route sends names meaning and factual claims and nothing else. No honorific logic exists in the codebase. The speech level you get back is whatever the model produced, and mixed input can yield mixed output or a silent unification in either direction.

The consequence is procedural rather than fatal. Ask for the register explicitly, then audit the sentence endings on the output. What you should not do is assume a subsystem is guarding it, because on this route there is not one.

The competing pages expose graded transformation: Simple, Balanced and Enhanced on one, Quick, Deep and Complete on another. This route exposes none of that. The request body carries route, family, input and instruction, and there is no fifth parameter.

What exists instead is a free-text field capped at 300 characters, appended verbatim to the system prompt. It is more expressive than a three-position slider and less predictable, because a slider maps to defined behaviour and a sentence maps to whatever the model does with it.

Treat it as the whole configuration of the run. Register, reader and an explicit hold-list of proper nouns is a good use of 300 characters. A vague adjective is a wasted one.

The category treats more transformation as strictly better, which is a reasonable position in English and a weaker one in Korean. Speech level encodes the relationship between writer and reader. Moving it does not make the text more human, it addresses it to a different person.

The second casualty is specificity. A rewrite that smooths rhythm can also smooth a hedge into an assertion, a change in factual strength dressed as a change in style. The instruction constrains claims, not confidence.

The practical defence is to bound the run rather than to trust the audit: shorter input, an explicit register, a named hold-list. Every page competing here sells a deeper transformation, and none of them discusses what one costs.

No, and the refusal is worth being specific about, because the competing evidence on this term is unusually confident. One page publishes a before-and-after table across four checkers and an aggregate bypass rate. Another advertises a rate above ninety-nine percent naming three products. A third lists ten detectors it claims to be built for.

Those are third-party outputs, measured by the seller, on their own text, against systems that change without notice. We have measured nothing, we are affiliated with no detector, and we will not predict what any of them will say about your text.

A detector reports observable signals, never proof of authorship, and it produces false results in both directions. The defensible claim is the narrow one: this rewrites Korean, and whether the result reads well is something you can verify yourself.

Five runs per rolling 24 hours, enforced server-side and counted from first use rather than reset on a calendar boundary. Input bounded at 12,000 characters, checked before transmission. Pasted text only, so document formats are out. No batch, no API, no presets, no history, no account.

No language detection. The route is Korean by URL, and the tool label the server builds is derived from the slug. Non-Korean input is rewritten in its own language without comment.

No deterministic spacing or spelling pass. 띄어쓰기 regularity is a real tell and not one this route is engineered to address. Spacing may shift as a side effect of rewriting, which is not the same as a spacing rule.