How Humanizer Prevents Pattern Detection False Positives: A Technical Deep Dive

Humanizer prevents pattern detection false positives through a layered defense system including multi-tell requirements, weak-alone flagging, contextual exclusions, voice-preserving overrides, and pattern hierarchies defined in SKILL.md.

The blader/humanizer repository implements a sophisticated AI-humanizing engine designed to rewrite AI-generated text into natural prose. Without proper constraints, this process risks creating false positives by flagging legitimate stylistic choices—such as em-dashes or bold headings—as artificial patterns. To prevent this, the tool employs specific safeguards detailed in its SKILL.md configuration file and detection logic.

Multi-Tell Requirements and Weak-Alone Detection

Humanizer's primary defense against over-correction relies on multi-tell verification and weak-alone flagging mechanisms defined in the skill documentation.

The "Several Tells Together" Rule

According to the SKILL.md section "When not to act", the engine requires multiple concurrent signals before triggering a rewrite. A single weak pattern—such as an isolated dash or bold label—remains ignored unless it appears alongside other tell-signals within the same passage. This "several tells together" rule ensures that isolated coincidences do not trigger unnecessary edits.

Weak-Alone Pattern Classification

Patterns deemed "weak alone" (including repeated dashes or decorative headings) carry explicit flags in the skill file under the "Weak alone" notes section. The engine treats these as low-confidence signals and mandates additional supporting patterns before initiating any transformation. This classification prevents the system from overreacting to common stylistic flourishes that appear in both human and AI writing.

Context-Aware Safeguards

Beyond pattern counting, Humanizer implements contextual exclusions that skip detection entirely in protected contexts.

Protected Context Blocks

As specified in SKILL.md under "When not to act", the system automatically bypasses detection inside:

• Quoted text passages
• Code blocks
• Headings that serve as titles
• Passages discussing the target phrase rather than using it directly

These exclusions prevent the model from flagging legitimate formatting or metalinguistic terminology.

Voice Preservation and Pattern Hierarchy

Humanizer balances sensitivity with specificity through adaptive voice matching and a tiered confidence system.

Writing Sample Adaptation

When users provide a writing sample, the engine adopts that sample's specific style characteristics, including dash frequency, bolding habits, and other stylistic quirks. As documented in the SKILL.md "Voice" section, this adaptive behavior prevents the skill from mistakenly labeling intentional authorial choices as AI-generated patterns. The system learns the user's baseline and adjusts its detection thresholds accordingly.

Pattern Strength Tiers

The skill implements a pattern hierarchy that orders detection rules by inherent strength. Patterns categorized in sections §1–§5 qualify as strong signals capable of triggering rewrites on single sightings. Conversely, patterns in sections §6–§20 register as weaker signals that require corroboration from stronger patterns before activation. This hierarchy, detailed under "Two rules follow from this" in SKILL.md, functions as a built-in confidence filter that reduces false-positive rates.

Implementation Example

The following Python implementation demonstrates how to invoke Humanizer while respecting its safeguard mechanisms, including voice preservation through sample submission:

import json
import requests

# Humanizer is delivered as a markdown skill; the consumer sends the text

def humanize(text: str, sample: str | None = None) -> str:
    payload = {"text": text}
    if sample:
        payload["sample"] = sample          # preserves author's voice

    resp = requests.post(
        "https://api.openai.com/v1/engines/humanizer/completions",
        json=payload,
        headers={"Authorization": f"Bearer YOUR_API_KEY"},
    )
    result = resp.json()
    return result["choices"][0]["message"]["content"]

# Example 1 – plain text (multiple tells trigger rewriting)

original = """
It's not just about the beat riding under the vocals; it's part of the aggression and atmosphere.
No aesthetic prior. No nostalgia.
"""

print(humanize(original))

# Example 2 – providing a sample that uses dashes heavily

sample = "The project — a daring experiment — succeeded."
print(humanize(original, sample=sample))

In the first call, only the "not-X-but-Y" contrast triggers rewriting because it meets the strong-pattern threshold. The second call respects the sample's dash usage, demonstrating how voice-preserving overrides prevent false-positive dash removal.

Summary

Humanizer prevents pattern detection false positives through five coordinated mechanisms:

• Multi-tell requirements mandating pattern corroboration before action
• Weak-alone flagging that labels low-confidence signals in SKILL.md
• Contextual exclusions protecting quotes, code blocks, and headings from detection
• Voice-preserving overrides that adapt to user-provided writing samples
• Pattern hierarchy tiers (§1–§5 strong vs. §6–§20 weak) that filter detection confidence

These safeguards work collectively within the SKILL.md architecture to balance detection sensitivity with correction specificity.

Frequently Asked Questions

What triggers a false positive in Humanizer?

A false positive typically occurs when the engine identifies a weak-alone pattern—such as a decorative dash or bold heading—in isolation without supporting context. Without the multi-tell requirement and pattern hierarchy safeguards, the system might incorrectly flag legitimate stylistic choices as AI-generated artifacts.

How does the pattern hierarchy prevent over-correction?

The pattern hierarchy restricts single-sight triggers to strong patterns (§1–§5), while requiring weaker patterns (§6–§20) to appear alongside stronger signals. This tiered system ensures that low-confidence detections cannot independently initiate rewrites, significantly reducing false-positive rates.

Can I customize which patterns Humanizer detects?

While the core pattern definitions reside in SKILL.md, you can influence detection behavior by submitting a writing sample when calling the API. This sample establishes a baseline for acceptable stylistic variations—including dash frequency and formatting choices—effectively customizing the false-positive threshold for your specific voice.

Where are the safeguard rules formally documented?

The primary safeguard definitions appear in SKILL.md within the blader/humanizer repository, specifically under the "When not to act", "Weak alone", "Voice", and "Two rules follow from this" sections. These sections collectively define the multi-tell requirements, contextual exclusions, and pattern strength classifications.

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