Methodology

How RageCheck detects manipulative patterns in content.

Overview

RageCheck uses a two-stage analysis pipeline: rule-based pattern detection followed by optional AI-powered contextual analysis. This hybrid approach balances speed, transparency, and accuracy.

The system analyzes text for linguistic patterns commonly associated with manipulative framing—language designed to provoke emotional reactions rather than inform. It does not assess factual accuracy or political bias.

Signal Categories

Content is analyzed across five distinct signal categories, each targeting specific manipulation patterns:

Loaded Language

Emotionally charged words designed to provoke reactions rather than convey information.

Examples: Dehumanizing terms, inflammatory adjectives, insults disguised as descriptors, words that presuppose guilt or malice.

Absolutist Phrasing

Black-and-white language that eliminates nuance and complexity.

Examples: "Always," "never," "everyone knows," "no one can deny," "the only way," "completely," "totally."

Threat & Panic

Fear-mongering language that emphasizes danger, catastrophe, or existential threats.

Examples: "Dangerous," "threat," "crisis," "catastrophe," "destroy," "end of," "collapse," urgent calls to action based on fear.

Us-vs-Them Framing

Tribal language that creates artificial divisions between groups.

Examples: "They want to," "those people," "the elite," "real Americans," "our side," "the enemy," collective blame attribution.

Engagement Bait

Phrases designed to maximize clicks, shares, and emotional responses.

Examples: "You won't believe," "share this before," "what happens next," "the truth about," "what they don't want you to know."

Scoring System

Rule-Based Detection

The first stage uses pattern matching against curated dictionaries of manipulative phrases. Each category has weighted terms—stronger manipulative signals receive higher weights.

Scores are normalized per 1,000 words to account for content length, ensuring short tweets and long articles are compared fairly.

AI Enhancement

When available, Claude AI reviews the rule-based findings to add context. This stage can adjust scores based on factors rules can't capture:

  • Distinguishing quotes from original statements
  • Recognizing academic or analytical discussion of extremism
  • Identifying satire or irony
  • Detecting manipulation tactics the rules missed

Score Interpretation

0-33
Low

Minimal manipulation signals

34-66
Medium

Some concerning patterns

67-100
High

Significant manipulation density

Limitations

  • Pattern detection is not perfect—false positives and negatives occur
  • Context matters: the same words can be manipulative or neutral depending on usage
  • Non-English content is not well supported
  • Very short content (under ~50 words) may produce unreliable scores
  • Sophisticated manipulation that avoids common patterns may score low
  • A high score does not mean content is false—just that it uses manipulative framing