Programmatic Text Humanizer & Pattern Disruptor

Paste your draft below to strip invisible Unicode watermarks, remove 60+ formulaic phrasing and participial-tail patterns, vary sentence rhythm for natural burstiness, and break 5-word k-gram chains—while keeping your facts, numbers, quotes, citations ([1], \cite{}), equations, and code 100% intact.

How programmatic pattern disruption works

Commercial "AI humanizers" route your text through another large language model. Because every language model selects tokens from a smooth probability distribution, LLM-rewritten text carries a fresh machine signature, frequently alters numbers or citations, and breaks LaTeX or code syntax.

This tool uses a 14-stage deterministic rule pipeline that targets the three mathematical layers measured by similarity scanners and statistical classifiers—without calling any generative model.

1. Layer 1: Lexical fingerprinting & the Winnowing 8-word guarantee

Document similarity engines (such as Turnitin and iThenticate) build sparse document fingerprints using the Winnowing algorithm (Schleimer, Wilkerson & Aiken, 2003):

1. Normalization: Text is lowercased and stripped of punctuation into a continuous word stream.
2. 5-gram slicing: A sliding window extracts contiguous sequences of $k = 5$ words.
3. Karp-Rabin rolling hash: Each 5-gram is mapped to a numeric hash in $O(1)$ time. Changing a single word inside any 5-word window produces an unrelated hash.
4. Window minimum selection ($w = 4$): The scanner moves a window of $w = 4$ hashes and records the minimum hash in each window. Any matching passage of $k + w - 1 = 8$ consecutive words is mathematically guaranteed to share at least one fingerprint.

2. Layer 2: Syntax diversity, burstiness & discourse rotation

Statistical AI classifiers measure two core distributions across a draft: perplexity (how predictable word choices are) and burstiness (the variance in sentence length and clause structure). Language models produce uniform 18-to-24-word sentences, participial `-ing` tails, and repetitive transition openers (Furthermore, Moreover, Additionally, In conclusion).

• PDTB 2.0 discourse rotation: Rotates formal connectors using the Penn Discourse Treebank taxonomy and tracks every replacement so the same transition never repeats in a paragraph.
• Participial tail & burstiness sculptor: Breaks trailing participial clauses (, thereby reducing..., , highlighting...) into standalone finite sentences and balances short and long sentences to raise the coefficient of variation ($\sigma / \mu$).

3. Layer 3: Zero-width watermark scrubbing & fact locking

Copied model outputs often carry invisible Unicode characters (`U+200B` zero-width space, `U+200C` non-joiner, `U+200D` joiner, `U+2060` word joiner, `U+FEFF` byte-order mark) and dense em-dashes (`—`). The scrubber strips these while locking LaTeX math ($...$), citations ([1], \cite{}), double-quoted strings ("..."), URLs, and numeric data inside placeholders before transformation.

Rule-based engine vs. LLM paraphrasers

Why non-native (ESL) writers get falsely flagged

A 2023 Stanford University study (Liang et al., Patterns) found that commercial AI detectors misclassified over 61% of TOEFL essays written by non-native English speakers as AI-generated. Because second-language writers often rely on standard transition words and steady sentence lengths, perplexity-based classifiers penalize clear, textbook prose. Adding natural sentence-length variety, plain verbs, and varied connectors helps restore a natural rhythm—especially when paired with our readability checker.

Programmatic humanizer FAQ

How does this humanizer rewrite text without using generative AI?

Generative paraphrasers feed your draft into another language model, which swaps one set of predictable token probabilities for another and often alters facts or citations. This engine is 100% deterministic and rule-based: it strips invisible zero-width watermarks, scrubs 60+ formulaic phrasing and participial-tail patterns, rotates discourse connectives, varies sentence length for natural burstiness, and breaks contiguous 5-word k-gram chains while locking numbers, quotes, citations and equations in place.

Does it protect quotes, LaTeX equations, citations and source code?

Yes. Before any linguistic rule runs, the placeholder masking layer automatically locks double-quoted passages ("..."), inline and display math ($...$, $$...$$), LaTeX commands (\cite{}, \ref{}, \begin{}), numeric and author-year citations ([1], (Author, 2024)), exact numbers, and 85+ scientific and compound terms. In Code Comments mode, a string-aware scanner rewrites only # and // comments and docstrings so executable code stays untouched.

What is Winnowing k-gram overlap and why does 8 words matter?

Document similarity scanners use the Winnowing algorithm (Schleimer, Wilkerson and Aiken, 2003) with k=5 word k-grams and a window of w=4 rolling hashes. By that math, any unchanged passage of k + w - 1 = 8 consecutive words is guaranteed to leave a matching fingerprint. Breaking 5-word chains and restructuring clauses changes the Karp-Rabin hashes across the sentence.

Is my text uploaded or stored on a server?

No. Every stage runs locally in your browser in plain JavaScript. Nothing is sent over the network, logged, or stored on any server.