TEST RECORD · ← all field reports

FIELD REPORTS/FIELD NOTE/ISSUE #129

New post

Most people can spot a chatbot email by now: the tidy three-part structure, the hedging, the sentence that announces what comes next. This model rewrites that voice out of a draft and leaves the facts alone, which is the part most rewriters get wrong.

MODELhumanizer
PUBLISHEDOctober 22, 2026
READ TIME3 min
CATEGORYFIELD NOTE

Field notes

01 What it does

You paste a draft that sounds machine-written, and it hands back the same message in a looser, more human voice. It cuts the hedging and the filler sentences, varies sentence length, and reorders paragraphs when that reads better. It works on emails, essays, reports and forum posts, in English and Chinese.

The difference from asking a chatbot to "make it sound human" is that this model was trained to keep every number, unit, date, name and quote exactly as written, and to add nothing. In the developer's test on 420 English rewrites, a strict checker found no factual change in 376, and most of the misses were a single word. On a separate set of 210 drafts, the AI detector Originality.ai, at its strictest setting, flagged 11 rewrites as AI.

It is a 12 billion parameter model fine-tuned from Google's Gemma 4, released under Apache 2.0, with a free desktop app for Mac and Windows that runs it offline.

02 How to try it

The demo Space runs the full model in your browser. Paste an AI-written email you have lying around, the kind that opens with "I hope this message finds you well", and run it. Then hold the rewrite next to the original and check the figures and names. If there is a phrase that must not change, like a product name or a quoted line, select it in the draft and click Create fact so it stays word for word. For anything private, download the desktop app instead: on first run it picks a model size that fits your computer's memory, down to 8 GB machines, and then works fully offline. It can import Word and PDF files too.

03 Caveat

The demo stores submitted text and rewrites anonymously to improve the model. It scrubs emails, phone numbers and ID numbers, but names written out in words are not caught, so use the desktop app for anything sensitive. It still changes a detail now and then, so read numbers, dates and names before you send. And where a school or employer has rules about AI help, those rules still apply.

04 What you can do with it

  • Loosen up a work email you drafted with a chatbot before you send it.
  • Rewrite a report or blog post so it stops sounding like a template, with the figures untouched.
  • Make a forum or Reddit answer read like a person typed it.
  • Run a long Word or Markdown document through the command-line tool, which keeps headings, tables and code intact.

Try the demo

View model page

Read this issue on neuralexpedition.com →