GPT-5.2 humanizer, limits stated

One request, one completion, returned as the rewritten text alone. The fixed instruction directs the model to accomplish the transformation implied by the tool name while preserving the original meaning and factual claims, and to return only the rewrite with no preamble, commentary, quotation marks or markdown fence.

Two parameters govern the experience. The rewriting branch runs at a markedly higher temperature than the analysis branch on this site, which is why the same passage returns differently worded rewrites across runs. And the completion carries a token ceiling set independently of the input limit.

Nothing else shapes a run except your passage and up to 300 characters of instruction. There is no style model, no learned voice, no stored profile, no mode.

The run is trivial. The specification and the comparison are the work.

Step 1

Submit a passage inside the 1 to 12,000 character band, and well inside it if the whole rewrite has to return.

Step 2

Attach falsifiable constraints in the 300 character instruction field: a sentence count, a paragraph cap, a voice, a protected element.

Step 3

Diff the output against the input. Semantic drift is the failure mode of this operation and it conceals itself inside better prose.

This route's traffic differs from the general humanizer term. The complaint attached to this version is shape: a published eight-task writing comparison found it returning what the testers called much less digestible, solid blocks of text where the preceding version used bolded subheads and bullet points.

That reframes the operation. Generic vocabulary can be substituted word by word. Structure cannot, which means the instruction carries most of the load on this route, and a request to sound more human will underperform a request to hold every paragraph to four sentences.

The vendor's own prompting guidance for this version reads as a specification for exactly that repair: avoid long narrative paragraphs, prefer compact bullets and short sections, break information into digestible chunks, use lists and tables where they help.

A constraint is useful when you can check whether it was met. These are drawn from OpenAI's published guidance for this model version and fit inside the 300 character field.

Numeric length

The guide's stated default is three to six sentences for a typical answer, and two sentences for a yes-or-no with a short explanation. A number is checkable; shorter is not.

Paragraph ceiling

A maximum sentence count per paragraph converts the structural complaint into an instruction with a pass or fail.

Lexical register

Simple language, short words, concrete verbs, active voice, all quoted directly from the same guidance.

Opening discipline

Be direct and start answering immediately, which removes the preamble sentence that most generated drafts open with.

Protected elements

Quotes verbatim, figures unchanged, sequence of points fixed. Unstated, all three are fair game for a rewrite optimising fluency.

The same guidance instructs the model to supply concrete examples, named entities, mechanisms, case examples and specific numbers or dates. That is a diagnosis of what flat writing lacks, and it exposes the boundary of this operation precisely.

A rewrite told to preserve meaning has access to one source of information, your text. If the draft names no entity, the rewrite names none. If it cites no figure, none appears. What returns is the same absence, better paced.

The practical consequence is a division of labour rather than a disappointment. Rhythm, padding, block structure and flatness are mechanical and the tool handles them. Specificity is editorial and remains yours, and it is usually the pass that decides whether the writing is worth reading.

Input and output ceilings are set separately. Input is validated against 12,000 characters, roughly two thousand words. The completion is bounded by a token limit corresponding to considerably less than that.

So a maximum-length submission cannot be guaranteed to return complete, and a truncated rewrite still consumes one of five runs. None of the competing pages in this result set publishes its completion budget, which is why none of them discusses the failure.

Submitting in units of two or three paragraphs avoids it entirely, and has the secondary benefit of showing you what the rewrite does to your voice before you commit the remaining runs.

The category's standard offer is detector evasion. Pages in this result set advertise up to 100 per cent human content scores, cite user counts in the tens of millions, and display before-and-after readings moving a sample from 1 per cent human to 99 per cent. Every one of those readings comes from a checker owned by the party doing the rewriting.

We publish no such number because we have measured nothing, and a prediction about a third-party classifier we do not control would be a guess wearing the costume of a feature. What is offered here is a rewrite you can read and judge in seconds.

The ethical limit is equally plain. Using a rewrite to conceal AI involvement in work represented as your own may constitute misconduct under the rules that apply to you, and no wording on any tool page alters that.

No modes, no tone presets, no readability settings, no uploaded writing sample, no persistent profile. One free-text field, one request at a time.

No file ingestion, no batch queue, no API, no account, no retained copy, and no undo, so keep your original. Usage is counted against a hash of the connection across a rolling 24 hour window.

It performs no cleaning of hidden or zero-width characters, which belongs to a separate deterministic family here that runs in the browser, and it holds no affiliation with OpenAI.