The Mathematics of Zero-AI Pattern Disruption

How do AI detectors like Turnitin AI, GPTZero, and CopyLeaks detect machine-written text? They do not "know" what was written by a human. Instead, they calculate two statistical variables: Perplexity (how predictable word sequences are) and Burstiness (the variation in sentence length and syntactic complexity).

1. Contiguous n-Gram Disruption

Detectors examine 3-gram and 4-gram contiguous sequences against precomputed probability tables. When common cliché patterns like "delve into the comprehensive tapestry" appear, the sequence probability spikes to near 100%. TURNITOUT's 4,068+ rule catalog targets contiguous n-grams, replacing predictable tokens with valid, lower-frequency synonyms and alternating voice without altering semantic meaning.

2. PDTB 2.0 Discourse Connective Rotation

AI models over-index on specific formal transition words: "Furthermore", "Moreover", "In conclusion", "Additionally". TURNITOUT rotates connectives using the Penn Discourse Treebank (PDTB 2.0) taxonomy, swapping rigid transitions with conversational or varied academic connectors to elevate natural sentence burstiness.

3. In-Place XML Run Parsing (.docx & .tex)

Traditional tools flatten text into plain ASCII, destroying formatting. TURNITOUT parses Word documents at the w:r (XML run) level. Text transformations are mapped directly into existing character nodes, keeping bold, italic, font family, color, hyperlinks, math matrices, and XML headers 100% intact.

4. AST String-Aware Comment Tokenizer

In programming code, natural language only exists inside comments and docstrings. TURNITOUT uses a character-by-character scanner that tracks string boundaries (e.g. print("# not a comment")) and isolates comments (#, //, /* */). All executable statements remain completely untouched, ensuring 100% executable safety.