OUR EDITORIAL APPROACH
How HumanizationAI edits your writing
Clarity, source fidelity, and a human final decision.
A useful suggestion should explain itself.
Our local scanner flags specific phrases and repeated openings with a reason. It is a rule-based writing aid. A sentence containing “furthermore” is not automatically bad; you decide whether the suggestion fits the context.
Rewrites are a separate, explicit action. Before starting, you see the selected paragraphs, the editing approach, and the exact input-word charge. A successful materially changed result consumes those words once. Internal checks and repair attempts do not multiply the charge.
A bounded editing pipeline.
The server freezes your selected source and preferences, reserves word credits, and queues a durable job. Exact protected spans are masked before the rewrite. We validate the returned schema, source references, protected details, and length. A separate model checks semantic fidelity.
The job has a normal 120-second hard deadline and at most four model invocations, including repair and verification. A failed, canceled-before-completion, blocked, or no-change job releases its reservation. A delivered result can be recovered without starting a new rewrite.
The model is an editor, not an authority.
Rewrites use OpenRouter with Google and OpenAI model routes. Data processing and retention also depend on those providers. We request routes that disallow provider data collection. Our application does not use customer drafts for cross-customer training.
We do not publish a human score, guarantee detector bypass, verify authorship, promise rankings, or claim universal factual accuracy. We also do not invent first-person anecdotes to make prose feel human.
What has been validated?
This release is an early product. Synthetic regression tests and operational checks are different from evidence that real writers prefer it. No comparative superiority or measured time-saving claim is made here.
The planned evaluation requires 20 permissioned real drafts, author and independent-editor review, and comparison with the author’s normal workflow and selected alternatives. Those research results are pending. Examples on this site are explicitly labeled synthetic or manually edited illustrations.