How Humanizer Identifies and Rewrites Borrowed Authority Claims (e.g., "Experts Believe")

Humanizer treats borrowed authority as Pattern 17 in its detection engine, using regex to flag phrases like "Experts believe" and either rewrite them with verified sources or strip the unsupported attribution while preserving the underlying fact.

The blader/humanizer repository provides an open-source skill for detecting and humanizing AI-generated text. When processing borrowed authority claims—vague attributions such as "Experts believe" or "Researchers say"—the tool applies specific rewrite rules that separate factual content from unsubstantiated prestige markers.

Pattern 17: The Core Detection Mechanism for Borrowed Authority

Humanizer categorizes borrowed authority as Pattern 17 within its tell-catalogue. According to the source code in [SKILL.md](https://github.com/blader/humanizer/blob/main/SKILL.md), this pattern targets specific linguistic signatures that signal unsupported authority boosts.

The Signature Patterns Triggering Detection

The regex defined in Pattern 17 captures three distinct authority markers:

Signature	Example
Explicit expert attribution	"Experts believe the new policy will reduce churn."
Prestige media citation	"Her views have been cited in The New York Times, BBC, Financial Times."
Unnamed authority references	"Researchers say this approach is optimal."

The Four-Stage Rewrite Pipeline

When handle_borrowed_authority() processes a sentence, it executes a strict preservation protocol to ensure factual accuracy while eliminating hollow authority markers.

Tell-Marking and Priority Sorting

During the initial tell-marking stage, Humanizer scans the entire document and flags every detected tell, ranking them by strength. Pattern 17 receives immediate priority when the text contains any of the signature phrases listed above.

Source Verification and Fallback Removal

The rewrite logic follows a deterministic branching path:

1. Fact-preservation rule: The underlying claim is considered sound; only the authority-boost requires modification.
2. Source verification: If the text contains a concrete, verifiable source (e.g., "According to Nature (2023)..."), Humanizer rewrites the sentence to cite that source directly.
3. Removal fallback: When no real source exists, the entire authority claim is stripped, leaving only the factual core to prevent hallucinated citations.

This logic is implemented in the skill's processing engine as follows:

def handle_borrowed_authority(sentence):
    if matches_pattern_17(sentence):
        source = extract_real_source(sentence)
        if source:
            return rewrite_with_source(sentence, source)
        else:
            return drop_authority_claim(sentence)   # keep the fact only

    return sentence

The matches_pattern_17 function executes the regex that catches phrases such as "Experts believe" and "cited in...".

Practical Implementation Examples

Direct Skill Invocation

When using the Humanizer skill in Claude or OpenAI agents, invoke it with the /humanizer command:

/humanizer

Experts believe that the new framework will cut latency by 40 %.

Result:

The new framework will cut latency by 40 %.

The vague attribution is removed while the specific metric remains intact.

Contextual Override with Writing Samples

You can provide a personal writing sample to guide the rewrite style:

/humanizer
Here’s a sample of my own writing:
> I usually back my claims with concrete data.

Now humanize this text:
> Experts claim the algorithm outperforms all baselines, as shown in many top conferences.

Result:

The algorithm outperforms the baselines, as shown in top-conference results.

The authority phrase "Experts claim" is excised, but the factual assertion about conference results persists.

Batch Processing Markdown Files

For document-wide processing, use file-mode execution:

humanizer docs/overview.md

If docs/overview.md contains:

"According to industry analysts, the platform will dominate the market."

Humanizer rewrites this to:

"Industry analysts predict the platform will dominate the market."

The vague "According to" framing is stripped, but the attribution remains because it references a concrete source category.

Key Source Files and Architecture

The borrowed authority detection system is defined across three primary files in the blader/humanizer repository:

• SKILL.md: Contains the master skill definition, including Pattern 17 regex specifications and the complete tell catalogue.
• README.md: Provides user-level documentation, installation instructions, and quick-start examples for invoking the authority-detection features.
• AGENTS.md: Defines metadata mappings that integrate the skill with various agent platforms including Claude and OpenAI.

Summary

• Humanizer identifies borrowed authority claims through Pattern 17, a regex-based detection system defined in SKILL.md.
• The tool distinguishes between concrete sources (which it preserves and clarifies) and vague attributions (which it removes).
• The handle_borrowed_authority() function implements a fact-preservation rule that never invents citations, ensuring the rewrite remains truthful.
• Users can invoke detection via direct skill calls, contextual writing samples, or batch file processing.

Frequently Asked Questions

What specific phrases trigger Pattern 17 detection?

Pattern 17 flags signatures such as "Experts believe," "Researchers say," "cited in [prestige publication]," and other unnamed authority markers. The exact regex is defined in SKILL.md and captures both explicit expert attributions and vague prestige citations.

Does Humanizer invent sources to replace removed authority claims?

No. According to the source code logic, if extract_real_source() returns no verifiable source, the drop_authority_claim() function executes a removal fallback that strips only the authority framing. The system never hallucinates citations or fabricates attribution to replace removed phrases.

How does Humanizer handle specific citations like "According to Nature (2023)"?

When the input contains a concrete, verifiable source such as a specific journal or publication date, Humanizer executes rewrite_with_source() to preserve that attribution. The tool restructures the sentence for clarity while maintaining the factual provenance, ensuring legitimate citations survive the humanization process.

Can I use Humanizer to process entire documents containing borrowed authority claims?

Yes. The repository supports batch processing via the CLI. Executing humanizer filename.md scans the entire document for Pattern 17 tells and applies the appropriate rewrite logic to each instance, making it suitable for cleaning up AI-generated reports or articles with pervasive authority markers.

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