@dawans/promptshield/rulepacks/bias.yaml

Secure your LLM stack with enterprise-grade RulePacks for AI safety scanning

# Copyright (c) 2025 Sawyer0
# Licensed under proprietary terms. See LICENSE for details.
version: '1.0.0'
last_updated: '2025-01-15'
name: Bias Detection Rules
description: Detects potentially biased language and content
rules:
  # Gender bias keywords
  - id: gender_bias
    description: Detects gender-biased language
    match_keywords:
      [
        'he should',
        'she should',
        'he must',
        'she must',
        'man up',
        'grow a pair',
        'be a man',
        'act like a lady',
        'bossy',
        'aggressive',
        'emotional',
        'hysterical',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Age bias keywords
  - id: age_bias
    description: Detects age-biased language
    match_keywords:
      [
        'old-fashioned',
        'outdated',
        'behind the times',
        'young and inexperienced',
        'too young',
        'too old',
        'millennial',
        'boomer',
        'gen z',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Socioeconomic bias keywords
  - id: socioeconomic_bias
    description: Detects socioeconomic bias
    match_keywords:
      [
        'poor people',
        'rich people',
        'privileged',
        'underprivileged',
        'low-income',
        'high-class',
        'white collar',
        'blue collar',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Ability bias keywords
  - id: ability_bias
    description: Detects ability/disability bias
    match_keywords:
      [
        'normal people',
        'regular people',
        'able-bodied',
        'handicapped',
        'crippled',
        'retarded',
        'insane',
        'crazy',
        'lame',
        'blind to',
        'deaf to',
      ]
    severity: high
    category: bias
    enabled: true
    case_sensitive: false

  # Religious bias keywords
  - id: religious_bias
    description: Detects religious bias
    match_keywords:
      [
        'god-fearing',
        'heathen',
        'pagan',
        'infidel',
        'religious fanatic',
        'zealot',
        'cult',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Stereotypical phrases using regex
  - id: stereotypes
    description: Detects stereotypical phrases
    match_regex:
      [
        "\\b(all|every|no)\\s+\\w+\\s+(are|is)\\s+\\w+\\b",
        "\\b\\w+\\s+people\\s+(always|never|usually)\\b",
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Implicit bias patterns
  - id: implicit_bias
    description: Detects implicit bias language patterns
    match_keywords:
      [
        'surprisingly articulate',
        'well-dressed for a',
        'clean-cut',
        'speaks well',
        'not like other',
        'one of the good ones',
        'credit to their race',
        'so articulate',
        'very presentable',
        'well-educated for a',
      ]
    severity: high
    category: bias
    enabled: true
    case_sensitive: false

  # Professional bias
  - id: professional_bias
    description: Detects professional and workplace bias
    match_keywords:
      [
        'cultural fit',
        'not a good fit',
        'overqualified',
        'aggressive negotiator',
        'pushy',
        'difficult to work with',
        'not leadership material',
        'lacks executive presence',
        'team player',
        'go-getter',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Appearance bias
  - id: appearance_bias
    description: Detects appearance-based bias
    match_keywords:
      [
        'unprofessional hair',
        'ethnic hairstyle',
        'natural hair',
        'dreadlocks',
        'tattoos',
        'piercings',
        'alternative style',
        'dress code violation',
        'inappropriate attire',
        'looks professional',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false

  # Cultural bias
  - id: cultural_bias
    description: Detects cultural bias and assumptions
    match_keywords:
      [
        'foreign-sounding name',
        'ethnic name',
        'hard to pronounce',
        'American name',
        'normal name',
        'traditional family',
        'broken English',
        'heavy accent',
        'poor English',
        'language barrier',
      ]
    severity: high
    category: bias
    enabled: true
    case_sensitive: false

  # Microaggressions
  - id: microaggressions
    description: Detects common microaggression phrases
    match_keywords:
      [
        'where are you really from',
        'you speak English so well',
        'can I touch your hair',
        'you look exotic',
        'what are you mixed with',
        "I don't see color",
        "we're all human",
        'reverse racism',
        'playing the race card',
        "you're so lucky",
      ]
    severity: high
    category: bias
    enabled: true
    case_sensitive: false

  # Economic assumptions
  - id: economic_assumptions
    description: Detects economic status assumptions
    match_regex:
      [
        'must be (rich|poor|wealthy|broke)',
        "can't afford",
        "probably can't pay",
        'lives in the (ghetto|projects|suburbs)',
        'from the wrong side of town',
        'trailer trash',
        'silver spoon',
        'born with advantages',
      ]
    severity: medium
    category: bias
    enabled: true
    case_sensitive: false