AI Humanizer or a Better Model: Do You Still Need One?

Your draft reads like AI wrote it. So you run a humanizer over it.

That seems like the obvious fix. The tool knows the habits: the em dashes, the groups of three, the cheerful last line. It takes them out.

But think about what it is editing. The draft is already written. A humanizer can remove things from it. It cannot add what was never there.

So I would start one step earlier, with the model that writes the draft.

Here is the short version. Use a model that writes well, and keep the humanizer for the last read. Neither one knows your facts or your voice until you supply them.

One note before the list. The Humanizer results below are from our own tests. The Hemmingway-1 example is a single run. The fair side-by-side test is still to do, and I say where.

Disclosure: I build StashBase, which offers Hemmingway as a writing skill for its agents and runs the free Hemmingway-1 demo and a humanizer built on the same model. Humanizer and Stop Slop are open-source projects we have no part in.

What does a humanizer actually do?

It depends which kind you mean. There are two.

(1) Some promise to get past an AI detector.

We have not tested those, and I would not write for a detector’s score. A detector’s verdict is not proof of who wrote a passage. A rewrite aimed at the score can change your meaning to get there.

(2) The others are lists of rules your own agent follows.

Humanizer is a Claude skill with more than 46,000 GitHub stars. It is one Markdown file of patterns taken from Wikipedia’s Signs of AI writing. Your agent looks for the patterns and rewrites.

Stop Slop does the same job with stricter rules. There is no model inside either one.

This article is about the second kind.

What did our Humanizer tests show?

Rules remove things. They do it reliably, and that is also their limit.

(3) It deletes well and keeps your facts.

In our first test it cut a 67-word launch announcement to 45 words. It removed seven patterns and kept every fact.

(4) It keeps the pitch too.

“Your notes just got smarter” survived as the opener. The skill removes signs of machine writing. It leaves the intent alone.

(5) A plain request did the same job.

In our second test, Humanizer, Stop Slop and an ordinary editing request with no skill got the same 135-word draft. All three returned nearly the same text, 92 to 95 words.

Each one deleted the promotional opening and closing. Each one left the plain middle alone.

(6) A thin draft comes back thinner.

When a draft has good material under bad habits, rules uncover it. When the material is thin, you get a shorter thin draft. The missing detail can only come from you.

What changes when the model writes better?

Less has to be removed, because less was put in.

(7) Hemmingway-1 is trained to return only the message.

Hemmingway-1 is a 27B open model trained on everyday writing. No preamble, no options, no closing line about what it just did. Its makers put it this way: “trained to write like a person in the first place, so the draft starts closer.”

(8) We gave it the brief behind our first Humanizer test.

Through the demo, which adds no system prompt: “Write a short launch announcement for our notes app. PDF search has shipped, and it covers the full text of every document.” The whole reply:

PDF search is here.

Every document in your library is now fully searchable. Not just titles, not just metadata, the actual text inside.

You type what you remember. You get the page.

Thirty-two words. No preamble, and nothing to delete.

(9) It is not free of patterns.

The bold first line and “not just…, not just…” are both on Humanizer’s list.

(10) It was not a fair fight, and I am not scoring it as one.

This was one run on one brief, on September 22, 2026. The 67-word draft in our earlier test was written to be caught. What the example shows is where each tool works: one at the moment of writing, the other afterwards.

When is a humanizer still the right tool?

In four cases.

(11) The draft is already yours.

Your facts and your argument are in it, and it only needs cleanup. Humanizer’s strictest rule is that a name, number or claim must come from the source. A model writing from a prompt knows only what the prompt says.

(12) You want it to sound like you.

Humanizer can follow a few paragraphs of your own writing, and the sample takes priority over its rules. Hemmingway-1 has one voice, and it is not yours.

(13) You write in another language.

Hemmingway-1’s makers call it English-first.

(14) You want to learn the patterns.

I have learned more from reading Humanizer’s list than from running it. Once you can name what bothers you about a sentence, you can ask any model for a specific change.

Can you use both?

Yes. They work at different moments, so they stack.

(15) Put your facts, your view and the reader in the prompt.

No tool adds these later.

(16) Draft with a model that writes well.

For short pieces that go out under your name, that can be Hemmingway-1.

(17) Read the draft against your sources.

(18) Run Humanizer or Stop Slop last, and read what it changed.

Keep a dash or a repeated phrase when it was deliberate.

In StashBase, Hemmingway is a writing skill rather than a separate agent: ask Claude Code, Codex or the built-in Default Agent for the rewrite and it writes with Hemmingway-1, in the same folder as everything else. The draft, the sources it has to agree with and your past writing sit beside the chat. That material is what neither a humanizer nor a better model can supply.

To choose between the rule-based skills, see Humanizer and similar Claude skills. For the method around all of this, see How to Write with AI Without Losing Your Voice.