Common False Positives Humanizer Is Designed to Ignore: Complete Pattern Reference

Humanizer explicitly ignores over 15 distinct categories of benign writing patterns—from perfect grammar and formal vocabulary to letter-style openings and em dashes—to avoid flagging legitimate human text as AI-generated.

The blader/humanizer repository provides a sophisticated skill for distinguishing AI-generated content from authentic human writing. Unlike simple detection tools that flag isolated stylistic choices, Humanizer implements a selective approach that recognizes what common false positives Humanizer is designed to ignore, ensuring that skilled human authors, technical writers, and formal documents remain untouched. Understanding these exclusions is critical for interpreting the tool’s rewrite decisions and trusting its output.

Stylistic and Grammatical Excellence

Humanizer treats high-quality writing as inherently human, explicitly excluding several markers of grammatical precision and vocabulary sophistication from its detection algorithms.

• Perfect grammar & consistent style: Professional editors and skilled writers routinely produce flawless prose. According to the source code analysis, Humanizer acknowledges that thorough proofreading and adherence to style guides like Chicago Manual do not indicate AI authorship.
• Formal or academic vocabulary: The presence of sophisticated terms such as "methodology" or "statistical inference" only triggers flags when appearing repetitively as part of a formulaic AI pattern. Isolated usage is considered normal human expression.
• "Bland" or "robotic" prose: Dry, neutral statements commonly found in technical documentation (e.g., "The system processes data and outputs results") are explicitly categorized as unreliable AI indicators. Many humans write plainly for clarity.

Structural Conventions and Formatting Markers

The tool preserves legitimate formatting artifacts and structural conventions that predate large language models or result from standard editing tools.

• Letter-style openings and closings: Traditional salutations such as "Dear Committee," or "Sincerely" are historical conventions that do not suggest machine generation.
• Curly quotes: Modern word processors automatically convert straight quotes to curly variants; this typographical feature alone carries no AI significance.
• Em dashes alone: While em dashes paired with formulaic sales rhythms may indicate AI-generated marketing copy, isolated usage for emphasis is recognized as standard editorial practice.
• Correct, complex formatting: Clean HTML markup, properly formatted tables, or template-generated structures are treated as neutral artifacts rather than machine signatures.

Rhetorical Devices and Stylistic Choices

Humanizer recognizes that deliberate stylistic techniques can superficially resemble AI patterns while serving legitimate rhetorical purposes.

• Mixed casual and formal registers: Authors frequently blend styles to suit specific audiences (e.g., "We’ve achieved great results, but further work is required"). This deliberate code-switching is preserved.
• One short sentence for emphasis: A single punchy line like "Success is inevitable" is respected as a stylistic choice; Humanizer only flags repeated sentence fragments as AI indicators.
• Deliberate repeated openings: Rhetorical devices such as anaphora (e.g., "She came. She saw. She conquered.") are distinguished from mechanical AI repetition.
• Transition words in isolation: Single connectors like "Additionally," "Moreover," or "Consequently" are harmless; the system only reacts to dense clusters of these transitions.
• Conversational markers: Words like "honestly" or "look" appearing mid-sentence are common in natural speech. They become flagged only when used as theatrical openers.

Content-Specific Legitimate Patterns

Certain content types contain characteristics that might appear artificial but serve necessary functions in human communication.

• Useful limits and disclaimers: Legal notices, safety warnings, and medical disclaimers are recognized as legitimate content requirements rather than AI artifacts.
• Real alternatives: Presenting realistic options that readers might actually choose (e.g., "You may use either JSON or XML") is preserved; only improbable, never-used alternatives are removed.
• Unsourced claims: The lack of citations for general statements (e.g., "Many experts believe...") is acknowledged as ubiquitous web practice, not exclusive to AI generation.
• Second-hand text: Phrases appearing inside quotations, titles, or examples are understood as being discussed rather than employed, and are left untouched by the rewriter.

Technical Implementation in SKILL.md

The complete taxonomy of exclusions is documented in the repository’s master skill definition. According to the blader/humanizer source code, these false positive rules are explicitly listed in SKILL.md between lines 95 and 115 under the section "What not to flag." This configuration file serves as the authoritative source for all pattern recognition logic, defining which indicators require multiple corroborating patterns before triggering a rewrite.

Humanizer’s architecture requires multiple simultaneous patterns before diagnosing a passage as AI-generated. A single indicator—regardless of how stereotypically "machine-like" it appears—is explicitly insufficient for flagging under the exclusion framework.

Practical Examples of False Positive Handling

The following examples demonstrate how Humanizer automatically applies these exclusions during processing:

{
  "skill": "humanizer",
  "input": "The system processes data and outputs results. It’s not just a tool, it’s a solution."
}

Result: Humanizer rewrites the second sentence (removing the "not just ..." pattern) while leaving the first sentence untouched. The bland technical description matches the false positive exclusion for neutral prose, while the formulaic construction in the second sentence triggers modification.

{
  "skill": "humanizer",
  "input": "Dear Committee,\n\nWe have achieved great results. The plan — though ambitious — was approved."
}

Result: The greeting "Dear Committee" and the em dash construction are preserved because they fall under the "letter-style opening" and "em dash alone" false positive exclusions. No rewrite occurs unless additional AI-specific patterns are detected elsewhere in the text.

Summary

• Humanizer maintains a whitelist of over 15 categories of benign writing patterns to prevent false accusations of AI authorship.
• Single stylistic indicators never trigger rewrites; the system requires multiple corroborating AI-specific patterns before intervening.
• Perfect grammar, formal vocabulary, standard punctuation (em dashes, curly quotes), and letter conventions are explicitly protected categories.
• The exclusion logic is defined in SKILL.md lines 95-115 and operates automatically without user configuration.
• Content inside quotations, legal disclaimers, and rhetorical devices like anaphora are recognized as human-authored and preserved.

Frequently Asked Questions

Does perfect grammar indicate AI-generated text to Humanizer?

No. Humanizer explicitly categorizes perfect grammar and consistent editorial style as false positives to ignore. The tool recognizes that skilled human writers, professional editors, and proofreaders routinely produce flawless prose, and this quality alone never triggers a rewrite.

Why does Humanizer preserve em dashes and curly quotes?

Em dashes and curly quotes are excluded because they typically result from standard editorial practices or word processor auto-formatting rather than AI generation. The source code treats isolated em dashes as legitimate emphasis tools, and curly quotes as typographical conversions that predate LLM usage. These only become suspicious when combined with specific formulaic rhythms.

How does Humanizer distinguish between rhetorical repetition and AI patterns?

Humanizer differentiates deliberate rhetorical devices like anaphora (e.g., "He laughed. He cried. He continued.") from mechanical AI repetition by context and pattern clustering. Single instances of repetition for emphasis are ignored as false positives, while repetitive structural patterns appearing multiple times throughout a text may trigger rewriting.

Where are the false positive rules documented in the repository?

The complete false positive exclusion list is documented in SKILL.md at lines 95-115, within the "What not to flag" section. This file in the blader/humanizer repository contains the master skill definition, pattern recognition rules, and the specific logic that prevents Humanizer from over-flagging legitimate human writing styles.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}