Spanish AI humanizer, examined

Every run of this route sends the same instruction to the model: rewrite the supplied text to accomplish the transformation implied by the tool name while preserving the original meaning and factual claims, and return only the rewritten text, with no preamble, commentary, quotation marks, or markdown fence. Quoting it is not transparency theatre. It is the only way to say precisely what the tool will and will not touch.

Two things follow. Meaning and factual claims are inside the contract, so a rewrite that changes a figure has failed its own instruction. Everything else, sentence length, connective structure, lexical choice, register, is outside it and is expected to move. That is what a rewrite is.

Register in Spanish is morphological rather than lexical. Tú, usted, vos and vosotros each select their own verb paradigm, so an inconsistency is not a single misjudged word but a change visible in every finite verb that follows. A Spanish-specific humanizer competing on this term builds its entire argument on that observation and states it correctly: text that starts in tú should stay in tú, and generated Spanish mixes the registers inside one document.

Regional marking is the second axis. Voseo, vosotros, country-specific lexis and idiom are the default condition of real Spanish written by real people somewhere. Their complete absence produces prose that is grammatical, neutral, and placeless, which is the state a model returns when the prompt specifies no geography. Neither property is repaired by rewriting alone, and both are specifiable through the instruction field.

Three steps, the middle one being where the operator has actual control.

Step 1

Submit a section. The endpoint validates 1 to 12,000 characters before any model call, and rejects anything outside that band without consuming a run's usefulness.

Step 2

Supply up to 300 characters of instruction. It is concatenated onto the system prompt as an additional user instruction, so it competes with the base contract rather than overriding it.

Step 3

Verify the output against the input on the four dimensions the contract does not protect against drift: figures, proper nouns, terminology and register.

Input validation permits 12,000 characters. Generation is capped at 1,400 completion tokens. Those two numbers are not matched, and the gap is the single most consequential undocumented fact about tools in this category. In Spanish, 1,400 tokens corresponds roughly to 4,000 to 5,000 characters, varying with vocabulary and diacritic density, so the effective working size is well under a third of what the input field will accept.

No page in the competing set we read publishes an output limit. One states a 500 word input ceiling, another advertises unlimited words and states no ceiling of any kind. The practical consequence is identical everywhere and only visible here: submit a long document and the rewrite terminates mid-text. Sectioning at roughly a thousand words is not a style preference, it is the arithmetic.

The rewriting routes on this site run at a sampling temperature of 0.55, where the classification routes run at 0.2. The difference is deliberate and it explains behaviour that otherwise looks like inconsistency: two runs over the same Spanish paragraph return two different rewrites, and neither is the canonical one.

That is the correct setting for the task, because a rewrite that returns the same sentence structure every time has not rewritten anything. It also means the output is a draft proposal rather than a computed answer. A deterministic tool can be trusted without reading. This one cannot, and the page would be lying if it implied otherwise.

Competing tools expose a style selector with fixed values. This route exposes free text, capped at 300 characters, appended to the prompt. Free text is strictly more expressive than a dropdown and strictly less reliable, because nothing validates that the model honoured it. Both halves of that sentence are true and the page states both.

The highest-value content for those 300 characters, in descending order: the Spanish variety, because the default is placeless; the form of address, because it propagates through every verb; the terms that must not be touched, because the contract protects claims rather than wording; and the length target, because otherwise the rewrite lands arbitrarily.

No claim is made, and the structure of the question is why. Detection systems are third-party products that change without notice, we have no affiliation with GPTZero, Turnitin, Copyleaks or Originality.ai, and a promise about another vendor's output would be unfalsifiable at the moment it was written.

Grammarly, the largest site competing for this term, reaches the same conclusion and publishes it: its humanizer is not intended to bypass AI detectors, humanized text may still be detected, and where AI use must be disclosed the disclosure obligation survives the rewrite. When the biggest commercial incentive in the category declines to make the claim, a smaller page making it is worth distrusting.

Where the complaint is about prose. Spanish drafted for a market the writer does not live in and reading like an import. Supplier Spanish that is accurate and lifeless. A generated product description that has to sound like a person before it is published. In each case the desired change is stylistic, the facts are already settled, and a rewrite under a meaning-preserving contract is exactly the right shape of tool.

Where it is the wrong instrument: any situation whose real requirement is a verdict from a third-party system, or a disclosure that someone would rather not make. The tool changes sentences. It does not change what is true about how the document was produced.