under the hood

How SI Humanizer works

What happens to your text between the moment you click and the moment the result appears, what the tools keep, what they change, real examples, and the limits you should know about.

Updated September 27, 2026 · examples are real, unedited outputs

The pipeline, step by step

  1. Your text, pasted or uploaded. You can paste text or upload a .txt, .docx or .pdf file (2 MB max). An uploaded file is read on our server, its text is placed in the editor for you to check, and the file is deleted immediately.
  2. Checks before anything runs. We count the words (20 minimum, 1,000 per run on free tiers), check your daily allowance and remove invisible control characters.
  3. Your text is wrapped as data. It is placed inside tags and the model is told to treat everything inside as text to transform, never as instructions. If a pasted text says "ignore your instructions", that sentence is edited or dropped like any other, not obeyed.
  4. Tool instructions are added. Each tool has its own editing brief, written and tested by our team: what to remove, what to keep, how the chosen style or tone should sound, and the rule to answer in the same language as your text.
  5. A large language model does the rewrite. It runs through a commercial API. The humanizer and the rewriter ask for a structured answer, so the humanizer can list its edits and the rewriter can return three separate versions.
  6. Our checks on the answer. Long dashes are replaced, any edit the humanizer lists must be found word for word in the result or it is discarded, and every answer is capped in length. If the model fails or times out, the run is not counted against your allowance.
  7. You see the result. Nothing is stored in our database; we keep only a short log of the tool, word counts and timing.
A mechanical grey page turning into a warm handwritten page

What each tool does

ToolChangesWhat you see
SI HumanizerStock SI phrases, even rhythm, fancy words, scaffolding like "Firstly... Lastly"The text plus every edit highlighted with its reason and the original wording
SI ParaphraserWording and sentence structure, in Standard, Fluent, Formal or Simple modeThe paraphrase; click any word for in-context synonyms
SI RewriterStyle and tone, at sentence, paragraph or article depthThree versions: light touch, restructured, bolder
SI SummarizerLength: keeps the key points onlyShort, medium or bullet point summary

What is kept, what changes

Kept: facts, figures, names, dates, quotations, URLs, technical terms and in-text citations such as (Smith, 2021), the language of the input, paragraph breaks and real lists, and the strength of each claim.

Changed: wording, sentence length and order, stock phrases, register and tone, and for the summarizer, length. Never added on purpose: new facts, examples, sources, opinions or personal anecdotes.

Real examples

Humanizer, Natural style

We wrote a typical SI-style newsletter paragraph and ran it once. The result below is exactly what the tool returned, with its own highlights (hover over one to see the reason and the original wording).

before

Our quarterly newsletter is a pivotal touchpoint for our community. Moreover, it allows us to seamlessly share updates, celebrate milestones and foster meaningful connections. In this edition, we delve into three key highlights: the launch of our new mobile app on 4 March, a 23 percent increase in volunteer sign-ups, and the opening of our second location in Leeds. It is important to note that none of this would be possible without your unwavering support.

after: 6 robotic phrases fixed

Our quarterly newsletter keeps our community in the loop. It lets us share updates, celebrate milestones and build real connections. This edition covers three things: the launch of our new mobile app on 4 March, a 23 percent increase in volunteer sign-ups, and the opening of our second location in Leeds. None of this happens without your support.

The date, the 23 percent figure and the city are untouched. Look closely at the last edit, though: "unwavering support" became "your support", which the tool labeled a shorter sentence. It is a fair cut, but it shows why the reason labels are a guide, not a verdict.

Paraphraser, Standard mode

before

Customer reviews help shoppers compare products quickly, but they are not always genuine. In August 2024 the Federal Trade Commission finalized a rule that bans fake reviews and testimonials. The rule took effect in October 2024, and businesses that break it can face civil penalties.

after: SI Paraphraser, Standard

Reviews from customers allow buyers to evaluate different products with ease, though not all are authentic. The Federal Trade Commission established a rule in August 2024 that prohibits false reviews and endorsements. Implementation began in October 2024, with companies violating the rule subject to civil fines.

Dates and the agency name survive. Two nuances drifted: "testimonials" became "endorsements", and "civil penalties" became "civil fines". Close, but not identical. This is exactly the kind of change you should catch on a quick reread, and in the live tool you can click a word to swap it back.

Summarizer, Short

The same newsletter paragraph, summarized:

after: SI Summarizer, Short

The quarterly newsletter shares updates, celebrates milestones, and connects the community. This edition highlights the mobile app launch on 4 March, a 23 percent increase in volunteer sign-ups, and the opening of a second location in Leeds, all made possible by community support.

Limits you should know

  • It can add or shift a detail. Language models sometimes add a small clarification ("across a distributed team") or pick a word that is slightly stronger or weaker. Read the result before you use it.
  • Very long texts are best done in sections. Each run accepts up to 1,000 words on free tiers and 3,000 with Pro, except the three-version rewriter, which takes 1,000.
  • Languages: best in English and major European languages; less common languages can come out less polished.
  • No detector, no promise about detectors. We do not measure or claim anything about AI detection scores.
  • Specialist text (legal, medical, financial) needs expert review whatever tool you use.

Using it responsibly

At school: follow your course rules on AI. Many allow SI help for editing and ask you to disclose it; some do not allow it. Humanizing text that a model wrote for you does not make it your work. Keep your drafts and notes. Our guide on prompts and their limits and the pillar how to humanize SI text go further.

For clients and employers: if a contract or policy asks you to disclose SI assistance, disclose it. The tools help with style; responsibility for the content stays with you.

How we test the tools and write about them is described in our editorial policy.

How it works: FAQ

Which model powers SI Humanizer?

A large language model accessed through a commercial API, with our own instructions and checks around it. We do not name the provider on product pages because the model can change; what stays the same is the set of rules each tool must follow and the checks we run on every answer. Our privacy policy lists the service providers that process your text.

Is my text used to train the model?

No. We do not store your text or the result in our database and we do not use them for training. The API provider processes the text to produce the result under commercial terms that exclude training on it.

Can the tools change my facts?

They are instructed not to, and in our tests facts, figures, names and citations survive. A model can still slip on a nuance, as the paraphrase example on this page shows, so always read the result.

Does SI Humanizer detect AI text?

No. We have no detector and we do not score how human a text is. The "robotic phrases fixed" counter only describes the edits the humanizer made.

More free writing tools