How Humanly works

The problem: AI writing has a fingerprint

AI-generated text is easy to spot once you know the tells. Sentence lengths barely vary — most land between fifteen and twenty-two words, over and over. Transitions come from a tiny stock pile: moreover, furthermore, in conclusion. Certain phrases appear at rates no human writer matches: things are constantly crucial, everything is a game-changer, and readers are forever invited to delve into topics. Ideas get packaged in threes. Claims get hedged into mush.

Detectors look for exactly these patterns, but so do human readers — a client, a professor, an editor. Even when nobody runs a detector, flat rhythm and stock phrasing make writing forgettable.

Pass one: forensic analysis (Pro)

On Pro, your text first goes through an analysis pass. The model reads it like a forensic editor and returns a structured list of specific problems: which AI-tell phrases appear and where, whether sentence rhythm is suspiciously uniform, which transitions are robotic, where the hedging piles up. This isn't shown as marketing fluff — it directly feeds the next step.

Pass two: the rewrite

The rewrite pass gets your text, your chosen strength and style, and (on Pro) the exact issue list from pass one. Its instructions are strict: vary sentence length aggressively, use natural transitions and contractions, cut the stock phrases entirely instead of swapping in synonyms — and preserve meaning, facts, names, numbers, and quotes exactly. Nothing is added, removed, or softened. The output stays in your input's language, whether that's English, French, or Arabic.

Strength levels

• Light — a subtle cleanup. Structure and paragraph order stay put; robotic wording gets fixed with the smallest possible touch.
• Medium — a noticeable rephrase. Sentence rhythm and flow get reworked, sentences merge or split where it reads better.
• Strong (Pro) — a full natural-voice redraft, as if a skilled writer rewrote the piece from the same notes. Only the meaning, facts, and order of ideas must survive.

Style presets

Five presets shape the voice: Neutral, Casual blog, Professional business, Academic-lite, and Marketing copy. Each is a short directive layered on top of the rewrite rules — the fact-preservation guarantee applies regardless of style.

Long texts

Pro accepts up to 3,000 words per request. Texts over about 1,200 words are split on paragraph boundaries, rewritten sequentially with the same style instructions so the voice stays consistent, then rejoined. You'll see per-part progress while it runs.

What the metrics mean

After every rewrite you get before/after numbers computed right in your browser (they cost nothing and your text never leaves the page for them): word and sentence counts, average sentence length, a Flesch reading-ease approximation, a count of known AI-tell phrases, and a naturalness meter built from sentence-length variance and tell density. The diff view highlights every word that changed; the AI-tells view marks what was flagged in your original.

Tips if you'd rather edit by hand

• Read your draft aloud. Anywhere you run out of breath at the same point in three sentences in a row, the rhythm is too even — shorten one, lengthen another.
• Search for moreover, furthermore, additionally, and in conclusion. Delete most of them; a paragraph break usually does the same job.
• Hunt the triads. If your draft keeps listing things in threes ("fast, reliable, and secure"), cut one item or restructure the sentence.
• Replace abstract intensifiers (crucial, vital, essential) with the concrete reason something matters.
• Add contractions where you'd use them in speech. "It's" and "don't" read as human; "it is" and "do not" read as generated.

Or paste your draft into the humanizer and let it do this in one pass — then use the diff view to learn what it changed.