An AI humanizer changes the surface. The risks live underneath.

A humanizer is a rewriting layer placed between a generated draft and its reader. It takes text and returns different text that is meant to read as less formulaic: shorter sentences, plainer vocabulary, fewer stock constructions.

The category is sold on two different promises, and they are worth separating. The first is editorial: make the prose less templated and easier to read. The second is adversarial: make a detector score lower. The first is a legitimate editing service. The second is a bet on somebody else's model staying unchanged.

Most products do both by accident, because the same edits produce both effects. That is exactly why you need to know what else the rewrite touched.

Almost every product is a wrapper around one of these four operations, sometimes with all four chained.

A humanizer that lowers a detector score has not established anything about authorship. It has changed the properties the detector measures, which is precisely what editing does. The score moves because the prose changed, not because the origin did.

The promise is also unstable. Detector vendors retrain on humanized output, humanizer vendors retrain against the new detectors, and the number you tested last quarter tells you nothing about this one. Building a publishing process on a moving target means re-testing forever.

There is a cheaper route to the same outcome. Cut the boilerplate, add the specifics, vary the rhythm. You get a lower score as a byproduct and a page that a reader will actually finish.

Rewriting is regeneration, and regeneration samples. Ask a model to paraphrase a sentence containing 14 percent and it may return around 15 percent, or simply fourteen. The sentence stays grammatical and confident, which is what makes the error expensive.

Numbers, names, dates, units, negations and qualifiers are the fragile parts. So are hedged claims: a cautious sentence that says a treatment may reduce symptoms can come back saying it reduces symptoms, and the medical meaning changes while the grammar stays perfect.

You will not catch this by reading for style. You catch it by diffing every numeral and proper noun against the source, and by re-reading the sentences with legal or clinical weight in them.

Synonym substitution is hostile to terminology. If your page says invoice, the humanizer may return bill; if it says churn, it may return attrition. On a product page that is a branding problem. On a help article it is a support ticket.

It can cost search visibility too. Pages rank for the phrases they use consistently, and a rewrite that replaces a target phrase with a synonym removes the phrase the page was written to answer.

Protect a list of terms before you run anything: product names, defined terms, headings, anchor text, the phrases you are targeting. Any tool that cannot be given a protected list is a tool you must proofread against the original, line by line.

Take a real paragraph of your own with three specific facts in it, run it through, and compare. You are looking for preservation first and improvement second.

Then check what it did to the structure. If it merged two paragraphs and dropped a caveat, you now know that every page it touches needs a structural review, which changes the economics completely.

Run the same paragraph twice. A tool that returns materially different text on each pass is regenerating rather than editing, so no output is reproducible and no review is reusable.

Finally, read the output for tone shift. A humanizer that makes a compliance page sound chatty has misunderstood the job.

Use the humanizer on prose only, never on a table, a list of specifications or a legal clause. Structure carries data, and data does not survive paraphrase.

Diff the output against the input before reading it for quality. Every changed numeral, name or qualifier is a decision, and you are the one who has to make it consciously.

Then edit by hand for the things a rewriting layer cannot know: this customer's actual terminology, what the team calls the feature internally, the number the support team quoted last week.

Keep the original and the rewritten version together in the record. If a claim is later disputed, the pair is the audit trail.

Sentence splitting. Long generated sentences broken into two shorter ones usually read better, and doing that by hand for a long page is tedious.

Connective removal. A tool that deletes furthermore and moreover is doing you a favour, provided it does not delete the logic they were standing in for.

Plain-language substitution for a consumer audience. Where a formal register was wrong to begin with, a tone shift earns its place.

Everything else, particularly vocabulary, is your call. The paragraph is the unit of editing, not the word.

If the draft has no specifics, a humanizer makes it read more smoothly while remaining empty, which is worse: boring prose is at least visibly boring, while confident prose with nothing behind it wastes the reader's time before they notice.

If the page carries regulated claims, a medical statement or a contractual obligation, write it yourself and have it reviewed by whoever owns the risk. Rewriting layers are not designed to preserve liability.

If you have notes, use the notes. Forty minutes with your own material beats an hour of checking what a rewriter changed and repairing it.

No humanizer adds knowledge. It can optimise the arrangement of what is already on the page, and that is the whole of its capability. A page that fails because it does not know the answer will fail just as much after being humanized.

That is why the useful sequence is always the same: gather the facts, write or edit the claim, vary the rhythm, then measure. A tool can accelerate the last two steps and cannot touch the first.

Used that way, a humanizer is a modest editing convenience. Used as a substitute for having something to say, it produces prose that reads confident and says nothing, which is the exact problem it was bought to solve.