Portuguese AI humanizer, untuned

A single server request carrying your text and one rewriting instruction: rewrite the supplied text to accomplish the transformation implied by the tool name, preserving the original meaning and factual claims, and return only the rewritten text without preamble, commentary, quotation marks or fencing.

Read that for what is not in it. There is no Portuguese in the instruction. No variant, no register, no vocabulary guidance, no orthographic convention. The only Portuguese-specific input the model receives is the tool's own name, interpolated into the sentence. Everything this tool does to Portuguese follows from a general-purpose rewriting request plus a label.

The competitive claim in this category is variant competence. One tool advertises models adjusted for Brazilian Portuguese with slang, idiom and informal register incorporated, promising writing like a real Brazilian. Another states it understands regional grammar, verb tenses, expressions and phrasing differences between the two variants, in both its marketing and its FAQ.

None of that is available here and none of it is asserted. There is no fine-tuned model in this codebase: the engine is a hosted general model identified by an environment variable. Saying so costs a marketing claim and buys something more durable, which is that every other statement on this page can be checked.

The control that does exist is a 300-character instruction field, appended to the system prompt verbatim as an additional user instruction. Stating a variant there is the mechanism, and it is the honest analogue of a dropdown.

Step two is not optional in practice. Absent an explicit variant, the output reflects whatever the model defaults to, which is not a property this page can specify for you.

Step 1

Paste the text within the enforced 1 to 12,000 character range.

Step 2

State variant, register and any protected terminology in the instruction field.

Step 3

Run once, then verify the variant against the two diagnostics below before reading for quality.

Since nothing enforces the instruction, verification is the reader's job, and two grammatical features settle it faster than reading for vocabulary.

Progressive aspect is the first. European Portuguese forms it periphrastically, estou a fazer, while Brazilian Portuguese uses the gerund, estou fazendo. The second is clitic placement: Brazilian Portuguese generally puts the object pronoun before the verb, European Portuguese generally after it. Both are frequent enough to appear in any paragraph of running prose, and both are binary rather than a matter of degree.

Inconsistency within a single output is the failure mode to watch for. A rewrite that mixes the two constructions has not chosen a variant, and no amount of vocabulary checking will show that as clearly.

Two classes of tell are commonly reported in Portuguese AI text, and a rewriting pass addresses them very unequally. The first is the connector stack, além disso, por outro lado, portanto, em resumo, at a density no unselfconscious writer sustains. That is a surface property and a rewrite genuinely can vary it.

The second is the contentless declarative, exemplified by sentences like este item é ideal para quem busca qualidade e durabilidade, or a inovação tecnológica transforma o futuro de maneira positiva. Each is well formed and asserts nothing contestable. A rewriting instruction that preserves meaning is constrained to preserve the absence of meaning too, so the output is a differently worded empty sentence. Those need deletion or an actual claim, and no rewriting tool can supply either.

Five runs per rolling 24 hours, counted from first use against a hash of the requesting address, with no account and no paid tier in the product. Input outside 1 to 12,000 characters is rejected before any request is made. Output is capped at 1,400 completion tokens, applied at the other end of the round trip and in a different unit, which is why a document can clear the input check and return truncated.

Sampling temperature is 0.55, so repeat runs are independent samples rather than refinements. Against a five-run ceiling this makes iteration expensive and front-loading the instruction cheap. For calibration: one Brazilian humanizer limits free processing to 500 words per run, and another places output limits behind a paid upgrade.

No claim is made and none should be believed from anyone. The published figures in this category are 97.5 percent bypass across Copyleaks, Turnitin, GPTZero and ZeroGPT on one tool, reported as 97.5 percent for Brazilian and 97.2 percent for European Portuguese, and a claim elsewhere of significantly reduced detection through manipulating perplexity and burstiness. That second tool then states that no tool guarantees a 100 percent result, which contradicts the framing above it.

What runs here is one rewriting request. Nothing targets a statistical property, no detector executes on our side, and no affiliation exists with any detection service. A percentage measured against software the claimant does not control, re-measured by nobody, is a marketing figure rather than a specification. For academic submission the governing document is the institution's policy, not a bypass rate.

Register conversion is the strongest case. Moving a formal Portuguese report toward a conversational article, or the reverse, is a transformation where the instruction to preserve meaning is exactly right and the output is easy to evaluate. Variant conversion is the second, provided the two diagnostics above are actually checked afterwards rather than assumed.

It earns least on text whose problem is that it has nothing to say. There the constraint that makes it safe, preserving meaning, is the same constraint that prevents it helping, and the correct intervention is editorial rather than mechanical. Knowing which of those two situations the draft is in determines whether a run is worth one of five.