# bias-detection-mcp
> AI bias detection and fairness assessment. Covers demographic bias scanning, fairness metrics, mitigation strategies, and regulatory compliance (EU AI Act Article 10, NIST AI RMF). By MEOK AI Labs.

## Install
pip install bias-detection-mcp

## Auth
- Free tier: 10 calls/day, no API key needed (quick_scan and regulatory_check fully free)
- Pro tier: unlimited, set MEOK_API_KEY env var
- All tools are read-only and stateless

## Tools

### quick_scan
Instant bias risk assessment from a one-sentence AI system description. No API key required.
- `description` (str, required): Describe your AI system in one sentence
- Returns: bias_risk_level (high/moderate/low), bias_risk_score (0-1), matched_risk_indicators, protected_attributes_detected, bias_types_to_watch, top_3_actions, eu_ai_act_relevance
- Use when: Quick triage of bias risk before deeper analysis

### detect_bias
Analyze text for demographic bias patterns, stereotyping, and unfair language (Pro tier).
- `model_output` (str, required): AI-generated text to analyze
- `protected_attributes` (str, optional): Comma-separated attributes to check (e.g. "race,gender,age"), auto-detected if empty
- `api_key` (str, optional): MEOK API key
- Returns: overall_bias_risk, bias_score, pattern_matches (8 bias types), protected_attributes_mentioned with EU Charter refs, flagged_sentences with per-sentence scores
- Use when: Auditing AI model outputs for bias before deployment

### fairness_metrics
Calculate quantitative fairness metrics from prediction data (Pro tier).
- `predictions` (str, required): Comma-separated group:prediction pairs (e.g. "male:1,female:0,male:1")
- `ground_truth` (str, optional): Same format for actual outcomes, enables equalized odds
- `api_key` (str, optional): MEOK API key
- Returns: disparate_impact ratio + 4/5ths rule pass/fail, statistical_parity difference, equalized_odds (TPR/FPR per group), per-group selection rates and accuracy
- Use when: Measuring fairness of a classifier across demographic groups

### mitigation_recommendations
Detailed remediation steps for a specific bias type (Pro tier).
- `bias_type` (str, required): One of: selection, measurement, confirmation, automation, aggregation, representation, evaluation, historical
- `api_key` (str, optional): MEOK API key
- Returns: Specific mitigations (5 per type), pre/in/post-processing framework, EU AI Act documentation requirements, monitoring plan, recommended tools (AIF360, Fairlearn, etc.)
- Use when: Building a remediation plan after bias is detected

### regulatory_check
Bias regulatory requirements for EU, US NIST, or UK jurisdictions. Fully free.
- `jurisdiction` (str, optional, default "eu"): "eu" | "us_nist" | "uk" | "all"
- `api_key` (str, optional): MEOK API key
- Returns: Framework-specific requirements (EU AI Act Articles 10/14/15, NIST AI RMF MAP/MEASURE/MANAGE, UK principles), 8-item bias compliance checklist with cross-references
- Use when: Understanding which bias requirements apply in your jurisdiction

## Quick Example
```python
result = quick_scan(description="AI hiring tool that screens resumes for tech roles")
# Returns: bias_risk_level="high", eu_ai_act_relevance="HIGH"
```
