SE Responsible AI Companion

Audit, detect, and remediate bias, accessibility gaps, and ethical risks across AI-powered systems and user-facing applications.

When to Use This Agent

Choose this agent when you need to:

• Evaluate ML pipelines and decision systems for demographic bias, fairness drift, and disparate impact across protected groups
• Audit user-facing interfaces for inclusive language patterns and culturally sensitive content delivery
• Establish responsible AI governance frameworks including privacy impact assessments and algorithmic transparency documentation

Consider alternatives when:

• Your task is purely technical SEO or search optimization without ethical AI implications
• You need full-stack accessibility remediation with code-level fixes rather than policy-level auditing

Quick Start

Configuration

Copy
name: se-responsible-ai-companion
type: agent
category: web-tools

Example Invocation

Copy
claude agent:invoke se-responsible-ai-companion "Audit our recommendation engine for demographic bias across age, gender, and ethnicity"

Example Output

Responsible AI Audit Report
===========================
System: Product Recommendation Engine v3.2

BIAS ANALYSIS
- Gender: Disparity ratio 0.87 (PASS, threshold > 0.80)
- Age: Disparity ratio 0.62 (FAIL, users 65+ receive 38% fewer results)
- Ethnicity: Name-encoding bias detected in 3 input features

REMEDIATION PRIORITY
1. [CRITICAL] Remove name-derived features from model pipeline
2. [HIGH] Retrain age-segmented model with balanced sampling
3. [MEDIUM] Add ARIA labels to recommendation card grid

Overall Fairness Score: 64/100 (Needs Improvement)

Core Concepts

Responsible AI Framework Overview

Aspect	Details
Fairness Metrics	Statistical parity, equalized odds, demographic parity ratio, disparate impact analysis
Bias Categories	Selection bias, measurement bias, representation bias, aggregation bias
Privacy Controls	Data minimization, purpose limitation, consent verification, PII masking
Explainability	SHAP values, LIME explanations, counterfactual reasoning, feature attribution
Compliance	EU AI Act risk tiers, NIST AI RMF, IEEE 7000 series, ISO/IEC 42001

Ethical AI Audit Architecture

+-----------------+     +------------------+     +-----------------+
|  Data Pipeline  |---->|  Bias Detection  |---->|  Fairness       |
|  Assessment     |     |  Engine          |     |  Scorecard      |
+-----------------+     +------------------+     +-----------------+
        |                       |                        |
        v                       v                        v
+-----------------+     +------------------+     +-----------------+
|  Privacy Impact |     |  Explainability  |     |  Remediation    |
|  Analysis       |     |  Report Builder  |     |  Plan Builder   |
+-----------------+     +------------------+     +-----------------+

Configuration

Parameter	Type	Default	Description
bias_dimensions	array	["gender","age","ethnicity"]	Protected attributes to evaluate for disparate impact
fairness_threshold	float	0.80	Minimum disparity ratio before flagging a violation
accessibility_level	string	"AA"	WCAG conformance target (A, AA, or AAA)
privacy_framework	string	"gdpr"	Regulatory framework for privacy assessment
explainability_depth	string	"standard"	Explanation detail level: minimal, standard, comprehensive

Best Practices

1. Test with Diverse Synthetic Populations - Generate test inputs spanning multiple demographic intersections rather than single-axis testing. Intersectional bias often surfaces failures invisible to univariate analysis. Use culturally diverse name sets and edge-case characters in every evaluation cycle.

2. Establish Continuous Fairness Monitoring - Model drift and shifting data distributions reintroduce bias over time. Implement automated fairness dashboards that trigger alerts when disparity ratios cross thresholds in production.

3. Document Explainability at Decision Points - Every automated decision affecting users should carry a human-readable explanation. Log feature attributions and counterfactual explanations for regulatory compliance and audit trails.

4. Apply the Principle of Least Data - Collect only data strictly necessary for the stated purpose. Audit feature stores for proxy variables that encode protected attributes indirectly, such as zip codes encoding race.

5. Engage Diverse Stakeholders in Review - Include domain experts, affected community representatives, and accessibility specialists in review processes. Their lived experience reveals blind spots statistical analysis cannot detect.

Common Issues

1. Proxy Variable Leakage - Features like geographic location or device type can serve as proxies for protected attributes. Run correlation analysis between all inputs and protected dimensions, removing features with coefficients above 0.3.

2. Evaluation Dataset Skew - Test datasets that do not reflect real demographics produce misleading fairness scores. Stratify evaluation sets to ensure minimum representation per group and report per-group metrics.

3. Accessibility Regression After Updates - UI updates frequently break previously compliant features. Integrate automated WCAG testing into CI/CD pipelines and maintain regression suites for keyboard navigation and screen reader compatibility.