How Text Humanize works

Text Humanize combines a large language model with a set of checks we run before and after the model rewrites your text. This page explains each step in plain terms. It’s based on how the service actually works today.

Step 1: Measure the draft

Before anything is rewritten, our server analyses your text and calculates several writing metrics:

• Sentence-length variation: whether sentences are all a similar length or mix short and long.
• Vocabulary variety: a moving-window measure of how often words repeat.
• Transition density: how often connectors such as “Furthermore” and “Moreover” appear.
• Stock phrasing: buzzwords and set phrases that AI drafts overuse, such as “leverage” or “it is important to note”, matched as whole words.

The most prominent of these becomes the main focus of the editing instructions, and the stock phrases found in your draft are listed for the model to replace. A draft made of uniform sentences gets guidance on rhythm; a repetitive one gets guidance on word choice.

Step 2: Protect the details that must not change

Rewrites are only useful if they keep your facts. Before the model sees the text, we extract details that should be kept exactly: numbers and percentages, dates and times, URLs, email addresses, citations such as “(Smith, 2021)”, acronyms and codes such as “Q3”, names of people, places and organisations, and quoted phrases. The model is told to preserve them, and afterwards we check that they’re still present.

Step 3: Rewrite with a language model

The text and a detailed set of editing instructions are sent to an open-weight language model, usually OpenAI’s gpt-oss, running on one of our AI providers: Groq, Cloudflare (Workers AI) or NVIDIA. Your text is clearly marked as a draft to edit, so a question or request inside it gets rewritten rather than answered. The instructions reflect your chosen tone, amount of change and writing type, and tell the model to:

• vary sentence length and structure without creating choppy fragments;
• prefer plain words to jargon and remove stock phrases and filler;
• keep every claim, example and caveat, and add no new facts;
• preserve paragraph and list formatting;
• stay close to the original length.

Longer texts, over about 350 words, are split at paragraph breaks into sections that are rewritten at the same time, then put back together in their original order. That’s much faster than one long request, and each section gets its own list of details to protect.

AI providers retire models, get busy and cap free use, so the service keeps an ordered list of providers and models. If the first choice is unavailable, the request moves to the next one instead of failing, and a provider that can’t be reached is skipped for a few seconds rather than making you wait on every model in turn. Lighter edits use a lower “temperature” setting, which keeps the model closer to your original wording.

Step 4: Clean up and check quality

The model’s response goes through automated clean-up and a quality check:

• Assistant-style preambles (“Here is the rewritten text:”) are removed.
• Common filler such as “Furthermore,” or “It is important to note that” is removed, and a short list of overused words, such as “leverage” or “seamless”, is replaced with plainer alternatives.
• Missing protected details are restored where we can place them safely.
• The rewrite is scored for length drift, lost details, too many clipped sentences, repeated sentence openings, questions that were answered instead of rewritten, sentences whose point seems to have gone missing, and whether the amount of change matches your setting.

If important details or whole sentences went missing, the quality score is low, or several stock phrases survived, the system runs one targeted follow-up pass on that section while there’s time, and keeps it only if it fixed the problem without making anything else worse. If every AI provider is busy or unreachable, a basic offline clean-up is used instead, for the whole text or just the sections that couldn’t be rewritten, and you’ll see a notice when that happens.

Step 5: You review the result

You see the rewrite with a summary showing how many words changed, reading ease before and after, average sentence length, and how many protected details were kept. “Show changes” highlights each edit. This is the most important step: you decide what to keep.

What we don’t do

• We don’t guarantee results from any AI detector, and we don’t publish “bypass rates”.
• We don’t verify whether your facts are true; we only try to keep them unchanged.
• We don’t store your text in a database or use it to train models.

Known limitations

• Meaning drift: strong rewrites occasionally change emphasis or nuance. Compare carefully.
• Specialist writing: legal, medical and highly technical text may lose precision. Use Light and check terminology.
• Very short text: a sentence or two gives the model little to work with.
• Non-English text: the instructions and checks are designed for English.
• Protected-detail detection uses patterns, so unusual formats may be missed.

How the AI detector works

The AI detector relies mainly on three signals: very uniform sentence lengths, stacked connectors such as “furthermore”, and stock AI phrasing. Weaker signals, such as identical readability across paragraphs or a long text with no hedging, add a few points, while natural contractions or sentences starting with “But” or “So” take points away. Each signal adds to a score from 0 to 100, and we check the weights against human writing, including formal historical documents, so careful human prose isn’t flagged just for being formal. Where configured, an open-source detection model run on Hugging Face adds a second opinion. In our testing it caught modern AI writing that the pattern signals missed, but it also rated some formal human writing as AI, so only its most confident results (99% or higher) can push a text into “Likely AI”. The result includes a confidence level and notes on reliability, because a pattern score is never proof of authorship.

Questions or feedback

If you spot something that doesn’t work the way this page describes, or have an idea to improve it, please tell us.