ChatGPT humanizer with claims stated

The humanizer operates on the distributional properties of model prose. Sentence lengths clustering near a mean, hedging constructions at high density, paragraph-initial formal connectives, tricolon enumeration arriving on a schedule, and a terminal paragraph that restates rather than concludes. The server prompt for this family instructs preservation of the original meaning and factual claims across that rewrite.

This route is a rewriter, which means the output is altered text. It produces no report, no score and no itemised list of what it changed. Diagnosis is a separate route on this site, and running it first is the difference between a targeted instruction and a general one.

The instruction is the only variable you control, and it dominates output quality.

Step 1

Submit between 1 and 12,000 characters of contiguous text, approximately 1,800 to 2,000 words.

Step 2

Supply up to 300 characters of constraint: named vocabulary to remove, named constructions to ban, and invariants to protect.

Step 3

Verify the output against the input on invariants: figures, dates, proper nouns, quoted passages, and qualifier strength.

Competing humanizer pages publish the vocabulary they say detectors attend to: multifaceted, leverage, paramount, ensure, facilitate, navigate, alongside the connective run of while, however and moreover. The list is real, and publishing it without a mechanism to act on it is the gap those pages leave open.

A 300-character instruction closes it, and the reason it works is worth being precise about. A request to make text sound human specifies no verifiable condition, so the rewrite has nothing to check itself against and falls back on a generic register shift. A named word is a condition with a truth value: either the output contains multifaceted or it does not. The same asymmetry applies to constructions. "Vary sentence length" is weak; "include at least three sentences under eight words" is checkable. Instructions that can be verified produce measurably more consistent results than instructions that cannot.

No, and we say so without qualification. We run no tests against Turnitin, GPTZero, Copyleaks or Originality.ai, tune nothing against their scoring, hold no affiliation with any of them, and make no prediction about what any of them will report.

The claims on this results page divide into three positions and the division is instructive. Grammarly, the most prominent operator in the category, states in its own FAQ that its humanizer is not intended to bypass AI detectors and encourages acknowledging AI assistance. A second competitor claims its service lowers scores on major detectors while conceding that no tool guarantees a pass, which is a claim and a disclaimer occupying the same paragraph. A tail of smaller domains simply asserts 100 per cent human text with no mechanism offered. Only the first position is one its publisher could defend if asked to show the working, which is the standard we hold ourselves to here.

The question has a governing answer and it is not ours to give: the written policy you are working under. Institutions and employers overwhelmingly frame these policies around disclosure of AI assistance rather than around which software processed a draft, which means the tool used is usually not the operative fact.

Underneath the single button, three distinguishable activities exist. Editing prose you wrote so it reads less uniformly is ordinary revision and requires no permission from anyone. Polishing assisted work you intend to disclose is permitted nearly everywhere, because the disclosure carries the obligation. Concealing generated work under a rule that prohibits it remains prohibited after any rewrite, since the rule addresses the origin of the work rather than its wording. The tool cannot distinguish these cases and does not try to, so the judgement sits with the person pressing the button, which is where it belongs.

Stated explicitly, because the category's marketing consistently implies otherwise.

An incorrect fact is carried across intact and expressed more fluently, which makes it marginally harder to catch on a reread. A missing citation is not supplied, because nothing is added that was not in the input. A weak argument becomes a smoothly written weak argument, since the rewrite addresses expression and has no access to the reasoning it expresses. And the uniformity that a humanizer removes is a symptom rather than the condition: prose generated without a specific thing to say converges on the same shape regardless of how its sentences are rearranged afterwards. Where that is the actual problem, the rewrite improves the reading experience and does not solve it.

Input runs from 1 to 12,000 characters, roughly 1,800 to 2,000 words, at 5 runs per rolling 24 hours measured from first use. A competing humanizer advertises capacity of approximately 5,000 words per call, and much of the long tail advertises unlimited free use. On raw capacity we are the smaller tool and the page states it rather than letting a user find out at paste time.

The constraint has a methodological edge worth naming. Rhythm is a property of a span, not of a sentence, so a rewrite evaluating sentence-length variance needs enough contiguous text for a distribution to exist. Splitting at section boundaries preserves that; splitting mid-paragraph destroys it and yields a flatter result than the same text produces in context. With 5 runs available, a long article split into three parts consumes most of a day's allowance, which makes diagnosing first and instructing precisely the economical order of operations rather than merely the tidy one.

Writers who drafted with a model, hold the argument themselves, and want the surface to stop signalling the drafting method. Professionals under a client or employer AI policy where the operative requirement is that published copy match an established house voice.

Teams producing at volume, where the convergence problem is acute: ten model-drafted pieces read as one piece, and readers register that long before any detection question arises. For that case the sequence that works is diagnosis first, then an instruction naming the specific recurring habits, then a rewrite. Running the rewrite blind on each piece reproduces the uniformity at one remove.