When Not to Run a Humanizer
Most visitors arrive here wanting the Humanize button. This page is for the times I tell people not to press it. A cleaner sentence is not always an honest sentence. REhume edits style. It does not change who is allowed to submit the work, and it does not make a detector’s percentage mean anything it did not mean before.
School, hiring, and written policies
If a course, journal, employer, or client forbids AI-assisted drafting, running a humanizer is not a workaround. It is still AI-assisted drafting, plus an extra step whose only purpose is to make that assistance harder to see. I will not pretend otherwise.
Read the actual rule. Some workplaces allow AI for outlines but require the prose to be yours. Some allow tools if you disclose them. Some ban detectors as evidence but still ban undisclosed generation. REhume cannot interpret your policy for you. If disclosure is required, put the disclosure in the document or the cover email. Do not ask a rewrite pass to stand in for that sentence.
I treat take-home exams, graded essays, and “write this in your own words” homework as off-limits unless the instructor has said tools are allowed. The fact that a detector might miss the result is irrelevant. Detectors are unreliable; that is a separate article. Unreliability is not permission.
Other people’s text
Do not paste a colleague’s email, a reporter’s article, or a vendor’s white paper and ship the output as yours. The tool does not check copyright or license. It will happily smooth someone else’s argument into something that sounds like a blog post. That is plagiarism with extra punctuation changes.
If you are editing with permission, copy desk, internal docs you own, a draft a coworker asked you to clean, say so in the workflow. Voice calibration is for your samples, not for mimicking a person who did not consent.
When the machine voice is the point
Chat transcripts, model eval notes, and “here is what the model said” appendices should keep chatbot artifacts. Stripping “As an AI language model” from a quote you are criticizing hides the evidence. Same for knowledge-cutoff disclaimers you are documenting as a failure mode.
If you are filing a bug about a model that inflates significance, keep the inflated sentence. Humanizing the exhibit makes the bug report worse.
Thin drafts that need facts, not polish
A paragraph that names no dates, no owners, and no numbers will come back shorter and still empty. REhume is instructed not to invent those things. If the source said “experts say” with no expert, the rewrite should not grow a fake citation. The honest fix is to look up the source or delete the claim.
I would rather you close the tab and open the spec, the spreadsheet, or the interview notes. Polish on top of a hole is how generic web pages get made. Google’s publisher rules call that low-value content. I call it a waste of the neuron budget.
Detector chasing
People paste the same paragraph ten times hunting for a green badge. That loop is the product I refuse to optimize. Scores move with model version, domain, and luck. A 12% “AI” result on a human recipe blog is a known failure. A 3% result on generated legal boilerplate is also a known failure.
If your actual goal is “pass GPTZero,” stop. If your goal is “a reader in my field would trust this paragraph,” read it aloud, check the proper nouns, and use the tool once. I wrote Why AI detectors are unreliable so I would not have to repeat this in email.
Live talks, jokes, and jargon
Humor dies in a “make it more natural” pass. So does tightly specified jargon that looks like a tell because models also use it. After a rewrite, restore the joke and the term of art. If you know you will restore most of the sentence, skip Humanize and edit the line yourself.
The manual editing pass exists for that mood: no API, no Turnstile, a printed checklist.
Sensitive material
Do not paste patient notes, unreleased financials, credentials, or anything your contract says must stay on-prem. REhume sends flagged windows to Workers AI. I do not keep the text after the response, but the inference still happens on Cloudflare’s infrastructure. Use a local editor for that class of document.
Rate limits and Turnstile are there so a scraper cannot burn the daily quota. They are not a privacy product.
What I still want you to use it for
Blog drafts you will fact-check. Internal RFCs you wrote. Product UI copy you will have a native speaker review. Emails that came out of a model with three “I hope this finds you well” openers. Those are in scope.
The test I use: would I be comfortable naming the tool if someone asked how the draft got from version one to version two? If the answer is no, do not run it.
REhume stays a small editor. It is not a conscience. That part is still yours.